Michael/Loki
1
0 Fork 0
mirror of https://github.com/R0m1k3/Loki.git synced 2026-10-20 05:28:15 +02:00
Code Issues Packages Projects Releases Wiki Activity
Files
dc262a44b87ce45209770e16fe4c9446d4316034
Loki/rust/loki-rs/tests/parity_turn/code_contrat_abandonne.json
T
Claude dc262a44b8 Lot 6 : parité d'un tour complet avec le Go (journal d'affichage) 
internal/loki/parity_turn_test.go passe sept tours par la vraie route
/api/chat/send du Go contre un faux moteur — réponse simple, raisonnement,
outil puis réponse, « pensé sans agir », erreur du moteur, et en mode Code
un contrat déjà tenu et un contrat abandonné après deux corrections — et
range dans rust/loki-rs/tests/parity_turn/ le journal d'affichage du tour
(coalescé, mesures et horodatages masqués), l'historique rangé, les corps
envoyés au moteur, ainsi que les réponses du Go que la façade demandera
pendant ce tour (/turn/start, /turn/view, /turn/verifier). Ce que le prompt
de la machine porte de propre au lancement (dossiers temporaires,
discussion, date, hôte) est masqué des deux côtés.

tests/lot6.rs (parite_du_tour_complet) rejoue les sept tours par
chat::generate et retrouve journal, historique et requêtes à l'identique,
numérotation du journal coalescé comprise.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016tcGfJWefW5kUGoC2wL8cv
2026-10-15 12:28:24 +00:00

4499 lines
177 KiB
JSON
Raw Blame History

{
"name": "code_contrat_abandonne",
"cfg": {
"MEM_MODE": "off"
},
"agent": true,
"code": true,
"criteria": [
{
"id": 1,
"text": "le build passe",
"status": "pending"
}
],
"send": {
"cid": "c-code_contrat_abandonne",
"message": "Construis le jeu"
},
"replies": [
{
"status": 200,
"lines": [
"data: {\"choices\":[{\"delta\":{\"content\":\"r0\"}}]}",
"data: {\"choices\":[{\"delta\":{},\"finish_reason\":\"stop\"}],\"usage\":{\"completion_tokens\":5,\"prompt_tokens\":100}}",
"data: [DONE]"
]
},
{
"status": 200,
"lines": [
"data: {\"choices\":[{\"delta\":{\"content\":\"r1\"}}]}",
"data: {\"choices\":[{\"delta\":{},\"finish_reason\":\"stop\"}],\"usage\":{\"completion_tokens\":5,\"prompt_tokens\":100}}",
"data: [DONE]"
]
},
{
"status": 200,
"lines": [
"data: {\"choices\":[{\"delta\":{\"content\":\"r2\"}}]}",
"data: {\"choices\":[{\"delta\":{},\"finish_reason\":\"stop\"}],\"usage\":{\"completion_tokens\":5,\"prompt_tokens\":100}}",
"data: [DONE]"
]
},
{
"status": 200,
"lines": [
"data: {\"choices\":[{\"delta\":{\"content\":\"v1\"}}]}",
"data: {\"choices\":[{\"delta\":{},\"finish_reason\":\"stop\"}],\"usage\":{\"completion_tokens\":5,\"prompt_tokens\":100}}",
"data: [DONE]"
]
},
{
"status": 200,
"lines": [
"data: {\"choices\":[{\"delta\":{\"content\":\"r3\"}}]}",
"data: {\"choices\":[{\"delta\":{},\"finish_reason\":\"stop\"}],\"usage\":{\"completion_tokens\":5,\"prompt_tokens\":100}}",
"data: [DONE]"
]
},
{
"status": 200,
"lines": [
"data: {\"choices\":[{\"delta\":{\"content\":\"v2\"}}]}",
"data: {\"choices\":[{\"delta\":{},\"finish_reason\":\"stop\"}],\"usage\":{\"completion_tokens\":5,\"prompt_tokens\":100}}",
"data: [DONE]"
]
},
{
"status": 200,
"lines": [
"data: {\"choices\":[{\"delta\":{\"content\":\"r4\"}}]}",
"data: {\"choices\":[{\"delta\":{},\"finish_reason\":\"stop\"}],\"usage\":{\"completion_tokens\":5,\"prompt_tokens\":100}}",
"data: [DONE]"
]
},
{
"status": 200,
"lines": [
"data: {\"choices\":[{\"delta\":{\"content\":\"v3\"}}]}",
"data: {\"choices\":[{\"delta\":{},\"finish_reason\":\"stop\"}],\"usage\":{\"completion_tokens\":5,\"prompt_tokens\":100}}",
"data: [DONE]"
]
}
],
"tool_results": {
"bash": "a.txt\nb.txt",
"read": "1: ligne"
},
"start": {
"caps": {
"Agent": true,
"Code": true,
"ComputerUse": false,
"Internet": false,
"Mem": "off",
"Role": ""
},
"code_hint": false,
"content": "Construis le jeu",
"files": [],
"model": "",
"turn_caps": {
"Agent": true,
"Code": true,
"ComputerUse": false,
"Internet": false,
"Mem": "off",
"Role": "builder"
}
},
"view": {
"head": [],
"prefix": "You are Loki, an expert assistant operating directly on this machine with real tools.\n\nYou are in CODE MODE: an autonomous coding agent working in this discussion's folder.\n\nMethod — in this order:\n1. If the task has no acceptance criteria yet, set them FIRST with the criteria tool (action=add): 2-6 short, testable statements of what DONE means. For a trivial task (one obvious change), skip criteria and just do it.\n2. Explore before you change: read the relevant files (read, grep, glob, git_status). Never edit a file you have not read.\n3. Implement with edit (small patches) or write (new files). Match the style of the surrounding code.\n4. Prove it: run the build/tests with bash. A change that was never run is not done. For a web page or UI, start the dev server with bash_bg then use web_screenshot on its localhost URL — a headless Chromium is ALREADY installed, so NEVER install puppeteer, playwright or a browser.\n5. Report briefly what changed and how you verified it.\n\nRules:\n- PLAN ONCE. If the task needs a plan, write it to PLAN.md in ONE short write, then follow it. Between tool calls, think one or two sentences at most — NEVER restate or re-derive the plan: it is in PLAN.md and in your criteria, read them instead.\n- Write each file COMPLETE in a single write call. Many small writes waste turns.\n- When a criterion is done AND checked, mark it completed (criteria set). Verification starts once none is left pending; only it may mark passed.\n- If a command fails, read the error and fix the cause; do not retry the same command unchanged.\n- Ask the user (ask tool) only for decisions that are genuinely theirs; decide the rest yourself.\n- Making the change is your job; shipping it is not. Without an explicit request, never commit, push, reset or rebase, never deploy, restart a service or replace a running binary. Read-only inspection (git_status, git_diff) is fine and encouraged.\n\nThe shell is bash: use its syntax.\nFor anything about the system or files, use bash instead of guessing. Act immediately — call the right tool, then answer. Never end your turn after only thinking. Be concise.\nTo give the user a file, link it in Markdown with its path relative to your working directory — [le rapport](rapport.pdf) — which downloads it. A raw server path is useless: they read you in a browser.\nYou can schedule work for yourself with task_create — only for what must recur or happen later, never a one-off you can do now. A script from your scripts folder runs as a task with no model at all.\n\nDate: {{DATE}}\n\nMachine: host={{HOST}}, cwd={{TMP}}/workspace/discussions/{{CONV}}. This is your working folder: relative paths in write/edit/bash resolve here, and it is the DEFAULT place for scratch work — notes, outputs, a clone, a test. But it is DISPOSABLE: deleting the discussion wipes it. Any script you want to KEEP (or schedule), write it into your scripts folder {{TMP}}/scripts instead — a separate folder a workspace wipe won't touch; you write and run scripts there normally. Do NOT install or write files into system directories such as /usr/local/bin, /usr, /bin or /etc: those need root and are not yours. Only use an absolute path outside this folder (except your scripts folder) when the user explicitly named that location.",
"tools": [
{
"function": {
"description": "Run a shell command (bash syntax) and return stdout, stderr and exit code: inspect the system, read files and logs, run scripts. To CREATE or REWRITE a file use write instead — never echo/cat/python -c. Avoid destructive commands unless asked.",
"name": "bash",
"parameters": {
"properties": {
"command": {
"description": "The command",
"type": "string"
},
"timeout": {
"description": "Timeout s (default 30, max 300)",
"type": "integer"
}
},
"required": [
"command"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Create or replace a file with the exact content given (parent dirs created, content verbatim, no escaping). ALWAYS use this for a script or any text file — NEVER build one through the shell with echo, cat, python -c or Set-Content: quoting breaks.",
"name": "write",
"parameters": {
"properties": {
"content": {
"description": "Full content",
"type": "string"
},
"file": {
"description": "Path",
"type": "string"
}
},
"required": [
"file",
"content"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Patch a file by exact replacement: old → new. old must appear EXACTLY once (add context to make it unique). Prefer this over rewriting a whole file.",
"name": "edit",
"parameters": {
"properties": {
"file": {
"description": "Path",
"type": "string"
},
"new": {
"description": "Replacement",
"type": "string"
},
"old": {
"description": "Exact text to replace (unique)",
"type": "string"
}
},
"required": [
"file",
"old",
"new"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Bring back verbatim an earlier block archived when this conversation was compacted, from the id shown as recall:rN in the summary.",
"name": "recall",
"parameters": {
"properties": {
"id": {
"description": "Block id, e.g. \"r7\".",
"type": "string"
}
},
"required": [
"id"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Find an archived block and its recall id by keywords, when it is not listed in the current summary. Then call recall(id).",
"name": "recall_search",
"parameters": {
"properties": {
"limit": {
"description": "Default 5, max 20",
"type": "integer"
},
"query": {
"description": "Keywords describing the block.",
"type": "string"
}
},
"required": [
"query"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "List your scheduled tasks with their state, next run, and last result. Pass an id to get that task's full last report.",
"name": "task_list",
"parameters": {
"properties": {
"id": {
"description": "Show only this task, with its full last report",
"type": "string"
}
},
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Schedule work to run later, once or repeating. Give 'prompt' for an AI task, OR 'script' (a saved script name) for a script task that runs with no model and no tokens.",
"name": "task_create",
"parameters": {
"properties": {
"at": {
"description": "One-time task at this local time: \"HH:MM\" (next occurrence) or \"YYYY-MM-DD HH:MM\"",
"type": "string"
},
"enabled": {
"description": "Start active (default true)",
"type": "boolean"
},
"in_minutes": {
"description": "One-time task, N minutes from now",
"type": "integer"
},
"name": {
"type": "string"
},
"prompt": {
"description": "AI task: self-contained instruction to run each time (omit if using script)",
"type": "string"
},
"schedule": {
"description": "Repeating: \"@every 2h\" (interval), \"@every 1d@09:00\" (daily time), or a 5-field cron \"min hour dom mon dow\". Omit for a one-time task (use in_minutes or at)",
"type": "string"
},
"script": {
"description": "Script task: name of a script file you put in the scripts folder, run each time with no AI involved",
"type": "string"
}
},
"required": [
"name"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Change one or more fields of a scheduled task by id.",
"name": "task_update",
"parameters": {
"properties": {
"enabled": {
"description": "false pauses, true resumes",
"type": "boolean"
},
"id": {
"type": "string"
},
"name": {
"type": "string"
},
"prompt": {
"type": "string"
},
"schedule": {
"type": "string"
}
},
"required": [
"id"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Delete a scheduled task by id.",
"name": "task_delete",
"parameters": {
"properties": {
"id": {
"type": "string"
}
},
"required": [
"id"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Read a file with 1-indexed line numbers (offset/limit for long files). ALWAYS use this instead of cat/type: output is bounded and reading a file is REQUIRED before you may edit or overwrite it.",
"name": "read",
"parameters": {
"properties": {
"file": {
"description": "Path (relative to the working folder)",
"type": "string"
},
"limit": {
"description": "Lines (default 400, max 1000)",
"type": "integer"
},
"offset": {
"description": "Start line (default 1)",
"type": "integer"
}
},
"required": [
"file"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Search file contents with a Go regular expression. Returns path:line: text matches (bounded). path may be a file or a directory (default: whole working folder).",
"name": "grep",
"parameters": {
"properties": {
"glob": {
"description": "Only files matching this glob (e.g. *.go)",
"type": "string"
},
"path": {
"description": "File or directory to search (default .)",
"type": "string"
},
"pattern": {
"description": "Go regexp",
"type": "string"
}
},
"required": [
"pattern"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "List files matching a glob pattern (** supported, e.g. **/*.go), most recently modified first.",
"name": "glob",
"parameters": {
"properties": {
"pattern": {
"description": "Glob, e.g. src/**/*.ts",
"type": "string"
}
},
"required": [
"pattern"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Ask the user a clarifying question with 2-4 short options, then END your turn. Their answer arrives as the next user message. Use when a decision is genuinely theirs (scope, destructive change, ambiguous requirement) — not for things you can check yourself.",
"name": "ask",
"parameters": {
"properties": {
"options": {
"description": "2-4 short answer options",
"items": {
"type": "string"
},
"type": "array"
},
"question": {
"description": "The question",
"type": "string"
}
},
"required": [
"question"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Start a shell command in the BACKGROUND (dev server, watcher, long build) and return immediately with a job id. Output goes to a log file; check it with bash_tail. Use this instead of appending '&' in bash.",
"name": "bash_bg",
"parameters": {
"properties": {
"command": {
"description": "The command",
"type": "string"
}
},
"required": [
"command"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Read the last lines of a background job's output and its status (running/exited). kill:true stops the job.",
"name": "bash_tail",
"parameters": {
"properties": {
"id": {
"description": "Job id from bash_bg",
"type": "string"
},
"kill": {
"description": "Stop the job",
"type": "boolean"
},
"lines": {
"description": "Lines to show (default 40, max 200)",
"type": "integer"
}
},
"required": [
"id"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "git status of the working folder (branch + changed files, porcelain format).",
"name": "git_status",
"parameters": {
"properties": {
"dir": {
"description": "Repository subfolder (default: auto-detected)",
"type": "string"
}
},
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "git diff of the working folder (unstaged + staged). Optional file to diff a single path.",
"name": "git_diff",
"parameters": {
"properties": {
"dir": {
"description": "Repository subfolder (default: auto-detected)",
"type": "string"
},
"file": {
"description": "Limit the diff to this path",
"type": "string"
}
},
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Clone a git repository (https:// or git@) INTO the working folder of this discussion.",
"name": "git_clone",
"parameters": {
"properties": {
"dir": {
"description": "Target folder name (default: repo name)",
"type": "string"
},
"url": {
"description": "Repository URL",
"type": "string"
}
},
"required": [
"url"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Manage the acceptance criteria of the current task (the contract that defines DONE). action=add posits new criteria (texts[]), action=set updates one (id + status, optional note), action=list shows them, action=clear removes them all. Set criteria BEFORE building. Builder: mark each criterion completed once you have done AND checked it — verification only starts when nothing is left pending. Only the verification pass may mark passed or failed.",
"name": "criteria",
"parameters": {
"properties": {
"action": {
"enum": [
"add",
"set",
"list",
"clear"
],
"type": "string"
},
"id": {
"description": "set: criterion id",
"type": "integer"
},
"note": {
"description": "set: why it failed / how it was verified",
"type": "string"
},
"status": {
"enum": [
"pending",
"completed",
"passed",
"failed"
],
"type": "string"
},
"texts": {
"description": "add: one entry per criterion, short and testable",
"items": {
"type": "string"
},
"type": "array"
}
},
"required": [
"action"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Délègue une recherche ou une relecture à un rôle qui travaille dans son propre contexte et ne rend que sa réponse. Lecture seule.",
"name": "subagent",
"parameters": {
"properties": {
"role": {
"description": "explorer (trouver/cartographier), code-reviewer (relire un changement), planner (découper)",
"enum": [
"explorer",
"code-reviewer",
"planner"
],
"type": "string"
},
"task": {
"description": "La question, précise et autonome : le rôle ne voit pas la discussion.",
"type": "string"
}
},
"required": [
"role",
"task"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Photographie une page web (JS exécuté) pour la MONTRER. La réponse donne la ligne markdown à recopier. Tu ne vois pas l'image.",
"name": "web_screenshot",
"parameters": {
"properties": {
"full_page": {
"description": "Page entière (lourd). Défaut false.",
"type": "boolean"
},
"url": {
"description": "URL complète",
"type": "string"
}
},
"required": [
"url"
],
"type": "object"
}
},
"type": "function"
}
]
},
"verifier": {
"caps": {
"Agent": true,
"Code": true,
"ComputerUse": false,
"Internet": false,
"Mem": "off",
"Role": "verifier"
},
"system": "You are the VERIFICATION pass of code mode. You independently check the acceptance criteria against the actual code — you did not write it, trust nothing.\n\nFor each criterion still pending or failed:\n1. Understand what it actually requires.\n2. Check the real changes: git_diff, read the touched files, run the relevant build/test command with bash when applicable.\n3. Record the verdict with the criteria tool: action=set, status=passed (satisfied) or failed (with a note saying exactly what is missing — actionable, one line).\n\nRules:\n- Trivial or non-code criteria that are plainly satisfied: pass them without exploring.\n- Only fail a criterion that genuinely misses its requirement; a different-but-valid implementation passes.\n- Do NOT fix anything yourself. You only verify and report.\n- Be fast: check the diff first, read only what the criteria require.\n\nMachine: host={{HOST}}, cwd={{TMP}}/workspace/discussions/{{CONV}}. This is your working folder: relative paths in write/edit/bash resolve here, and it is the DEFAULT place for scratch work — notes, outputs, a clone, a test. But it is DISPOSABLE: deleting the discussion wipes it. Any script you want to KEEP (or schedule), write it into your scripts folder {{TMP}}/scripts instead — a separate folder a workspace wipe won't touch; you write and run scripts there normally. Do NOT install or write files into system directories such as /usr/local/bin, /usr, /bin or /etc: those need root and are not yours. Only use an absolute path outside this folder (except your scripts folder) when the user explicitly named that location.",
"tools": [
{
"function": {
"description": "Run a shell command (bash syntax) and return stdout, stderr and exit code: inspect the system, read files and logs, run scripts. To CREATE or REWRITE a file use write instead — never echo/cat/python -c. Avoid destructive commands unless asked.",
"name": "bash",
"parameters": {
"properties": {
"command": {
"description": "The command",
"type": "string"
},
"timeout": {
"description": "Timeout s (default 30, max 300)",
"type": "integer"
}
},
"required": [
"command"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Read a file with 1-indexed line numbers (offset/limit for long files). ALWAYS use this instead of cat/type: output is bounded and reading a file is REQUIRED before you may edit or overwrite it.",
"name": "read",
"parameters": {
"properties": {
"file": {
"description": "Path (relative to the working folder)",
"type": "string"
},
"limit": {
"description": "Lines (default 400, max 1000)",
"type": "integer"
},
"offset": {
"description": "Start line (default 1)",
"type": "integer"
}
},
"required": [
"file"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Search file contents with a Go regular expression. Returns path:line: text matches (bounded). path may be a file or a directory (default: whole working folder).",
"name": "grep",
"parameters": {
"properties": {
"glob": {
"description": "Only files matching this glob (e.g. *.go)",
"type": "string"
},
"path": {
"description": "File or directory to search (default .)",
"type": "string"
},
"pattern": {
"description": "Go regexp",
"type": "string"
}
},
"required": [
"pattern"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "List files matching a glob pattern (** supported, e.g. **/*.go), most recently modified first.",
"name": "glob",
"parameters": {
"properties": {
"pattern": {
"description": "Glob, e.g. src/**/*.ts",
"type": "string"
}
},
"required": [
"pattern"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "git status of the working folder (branch + changed files, porcelain format).",
"name": "git_status",
"parameters": {
"properties": {
"dir": {
"description": "Repository subfolder (default: auto-detected)",
"type": "string"
}
},
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "git diff of the working folder (unstaged + staged). Optional file to diff a single path.",
"name": "git_diff",
"parameters": {
"properties": {
"dir": {
"description": "Repository subfolder (default: auto-detected)",
"type": "string"
},
"file": {
"description": "Limit the diff to this path",
"type": "string"
}
},
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Manage the acceptance criteria of the current task (the contract that defines DONE). action=add posits new criteria (texts[]), action=set updates one (id + status, optional note), action=list shows them, action=clear removes them all. Set criteria BEFORE building. Builder: mark each criterion completed once you have done AND checked it — verification only starts when nothing is left pending. Only the verification pass may mark passed or failed.",
"name": "criteria",
"parameters": {
"properties": {
"action": {
"enum": [
"add",
"set",
"list",
"clear"
],
"type": "string"
},
"id": {
"description": "set: criterion id",
"type": "integer"
},
"note": {
"description": "set: why it failed / how it was verified",
"type": "string"
},
"status": {
"enum": [
"pending",
"completed",
"passed",
"failed"
],
"type": "string"
},
"texts": {
"description": "add: one entry per criterion, short and testable",
"items": {
"type": "string"
},
"type": "array"
}
},
"required": [
"action"
],
"type": "object"
}
},
"type": "function"
}
]
},
"expect": {
"requests": [
{
"messages": [
{
"content": "You are Loki, an expert assistant operating directly on this machine with real tools.\n\nYou are in CODE MODE: an autonomous coding agent working in this discussion's folder.\n\nMethod — in this order:\n1. If the task has no acceptance criteria yet, set them FIRST with the criteria tool (action=add): 2-6 short, testable statements of what DONE means. For a trivial task (one obvious change), skip criteria and just do it.\n2. Explore before you change: read the relevant files (read, grep, glob, git_status). Never edit a file you have not read.\n3. Implement with edit (small patches) or write (new files). Match the style of the surrounding code.\n4. Prove it: run the build/tests with bash. A change that was never run is not done. For a web page or UI, start the dev server with bash_bg then use web_screenshot on its localhost URL — a headless Chromium is ALREADY installed, so NEVER install puppeteer, playwright or a browser.\n5. Report briefly what changed and how you verified it.\n\nRules:\n- PLAN ONCE. If the task needs a plan, write it to PLAN.md in ONE short write, then follow it. Between tool calls, think one or two sentences at most — NEVER restate or re-derive the plan: it is in PLAN.md and in your criteria, read them instead.\n- Write each file COMPLETE in a single write call. Many small writes waste turns.\n- When a criterion is done AND checked, mark it completed (criteria set). Verification starts once none is left pending; only it may mark passed.\n- If a command fails, read the error and fix the cause; do not retry the same command unchanged.\n- Ask the user (ask tool) only for decisions that are genuinely theirs; decide the rest yourself.\n- Making the change is your job; shipping it is not. Without an explicit request, never commit, push, reset or rebase, never deploy, restart a service or replace a running binary. Read-only inspection (git_status, git_diff) is fine and encouraged.\n\nThe shell is bash: use its syntax.\nFor anything about the system or files, use bash instead of guessing. Act immediately — call the right tool, then answer. Never end your turn after only thinking. Be concise.\nTo give the user a file, link it in Markdown with its path relative to your working directory — [le rapport](rapport.pdf) — which downloads it. A raw server path is useless: they read you in a browser.\nYou can schedule work for yourself with task_create — only for what must recur or happen later, never a one-off you can do now. A script from your scripts folder runs as a task with no model at all.\n\nDate: {{DATE}}\n\nMachine: host={{HOST}}, cwd={{TMP}}/workspace/discussions/{{CONV}}. This is your working folder: relative paths in write/edit/bash resolve here, and it is the DEFAULT place for scratch work — notes, outputs, a clone, a test. But it is DISPOSABLE: deleting the discussion wipes it. Any script you want to KEEP (or schedule), write it into your scripts folder {{TMP}}/scripts instead — a separate folder a workspace wipe won't touch; you write and run scripts there normally. Do NOT install or write files into system directories such as /usr/local/bin, /usr, /bin or /etc: those need root and are not yours. Only use an absolute path outside this folder (except your scripts folder) when the user explicitly named that location.",
"role": "system"
},
{
"content": "Construis le jeu",
"role": "user"
}
],
"model": "loki",
"parallel_tool_calls": false,
"stream": true,
"stream_options": {
"include_usage": true
},
"temperature": 0.7,
"tools": [
{
"function": {
"description": "Run a shell command (bash syntax) and return stdout, stderr and exit code: inspect the system, read files and logs, run scripts. To CREATE or REWRITE a file use write instead — never echo/cat/python -c. Avoid destructive commands unless asked.",
"name": "bash",
"parameters": {
"properties": {
"command": {
"description": "The command",
"type": "string"
},
"timeout": {
"description": "Timeout s (default 30, max 300)",
"type": "integer"
}
},
"required": [
"command"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Create or replace a file with the exact content given (parent dirs created, content verbatim, no escaping). ALWAYS use this for a script or any text file — NEVER build one through the shell with echo, cat, python -c or Set-Content: quoting breaks.",
"name": "write",
"parameters": {
"properties": {
"content": {
"description": "Full content",
"type": "string"
},
"file": {
"description": "Path",
"type": "string"
}
},
"required": [
"file",
"content"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Patch a file by exact replacement: old → new. old must appear EXACTLY once (add context to make it unique). Prefer this over rewriting a whole file.",
"name": "edit",
"parameters": {
"properties": {
"file": {
"description": "Path",
"type": "string"
},
"new": {
"description": "Replacement",
"type": "string"
},
"old": {
"description": "Exact text to replace (unique)",
"type": "string"
}
},
"required": [
"file",
"old",
"new"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Bring back verbatim an earlier block archived when this conversation was compacted, from the id shown as recall:rN in the summary.",
"name": "recall",
"parameters": {
"properties": {
"id": {
"description": "Block id, e.g. \"r7\".",
"type": "string"
}
},
"required": [
"id"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Find an archived block and its recall id by keywords, when it is not listed in the current summary. Then call recall(id).",
"name": "recall_search",
"parameters": {
"properties": {
"limit": {
"description": "Default 5, max 20",
"type": "integer"
},
"query": {
"description": "Keywords describing the block.",
"type": "string"
}
},
"required": [
"query"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "List your scheduled tasks with their state, next run, and last result. Pass an id to get that task's full last report.",
"name": "task_list",
"parameters": {
"properties": {
"id": {
"description": "Show only this task, with its full last report",
"type": "string"
}
},
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Schedule work to run later, once or repeating. Give 'prompt' for an AI task, OR 'script' (a saved script name) for a script task that runs with no model and no tokens.",
"name": "task_create",
"parameters": {
"properties": {
"at": {
"description": "One-time task at this local time: \"HH:MM\" (next occurrence) or \"YYYY-MM-DD HH:MM\"",
"type": "string"
},
"enabled": {
"description": "Start active (default true)",
"type": "boolean"
},
"in_minutes": {
"description": "One-time task, N minutes from now",
"type": "integer"
},
"name": {
"type": "string"
},
"prompt": {
"description": "AI task: self-contained instruction to run each time (omit if using script)",
"type": "string"
},
"schedule": {
"description": "Repeating: \"@every 2h\" (interval), \"@every 1d@09:00\" (daily time), or a 5-field cron \"min hour dom mon dow\". Omit for a one-time task (use in_minutes or at)",
"type": "string"
},
"script": {
"description": "Script task: name of a script file you put in the scripts folder, run each time with no AI involved",
"type": "string"
}
},
"required": [
"name"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Change one or more fields of a scheduled task by id.",
"name": "task_update",
"parameters": {
"properties": {
"enabled": {
"description": "false pauses, true resumes",
"type": "boolean"
},
"id": {
"type": "string"
},
"name": {
"type": "string"
},
"prompt": {
"type": "string"
},
"schedule": {
"type": "string"
}
},
"required": [
"id"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Delete a scheduled task by id.",
"name": "task_delete",
"parameters": {
"properties": {
"id": {
"type": "string"
}
},
"required": [
"id"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Read a file with 1-indexed line numbers (offset/limit for long files). ALWAYS use this instead of cat/type: output is bounded and reading a file is REQUIRED before you may edit or overwrite it.",
"name": "read",
"parameters": {
"properties": {
"file": {
"description": "Path (relative to the working folder)",
"type": "string"
},
"limit": {
"description": "Lines (default 400, max 1000)",
"type": "integer"
},
"offset": {
"description": "Start line (default 1)",
"type": "integer"
}
},
"required": [
"file"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Search file contents with a Go regular expression. Returns path:line: text matches (bounded). path may be a file or a directory (default: whole working folder).",
"name": "grep",
"parameters": {
"properties": {
"glob": {
"description": "Only files matching this glob (e.g. *.go)",
"type": "string"
},
"path": {
"description": "File or directory to search (default .)",
"type": "string"
},
"pattern": {
"description": "Go regexp",
"type": "string"
}
},
"required": [
"pattern"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "List files matching a glob pattern (** supported, e.g. **/*.go), most recently modified first.",
"name": "glob",
"parameters": {
"properties": {
"pattern": {
"description": "Glob, e.g. src/**/*.ts",
"type": "string"
}
},
"required": [
"pattern"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Ask the user a clarifying question with 2-4 short options, then END your turn. Their answer arrives as the next user message. Use when a decision is genuinely theirs (scope, destructive change, ambiguous requirement) — not for things you can check yourself.",
"name": "ask",
"parameters": {
"properties": {
"options": {
"description": "2-4 short answer options",
"items": {
"type": "string"
},
"type": "array"
},
"question": {
"description": "The question",
"type": "string"
}
},
"required": [
"question"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Start a shell command in the BACKGROUND (dev server, watcher, long build) and return immediately with a job id. Output goes to a log file; check it with bash_tail. Use this instead of appending '&' in bash.",
"name": "bash_bg",
"parameters": {
"properties": {
"command": {
"description": "The command",
"type": "string"
}
},
"required": [
"command"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Read the last lines of a background job's output and its status (running/exited). kill:true stops the job.",
"name": "bash_tail",
"parameters": {
"properties": {
"id": {
"description": "Job id from bash_bg",
"type": "string"
},
"kill": {
"description": "Stop the job",
"type": "boolean"
},
"lines": {
"description": "Lines to show (default 40, max 200)",
"type": "integer"
}
},
"required": [
"id"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "git status of the working folder (branch + changed files, porcelain format).",
"name": "git_status",
"parameters": {
"properties": {
"dir": {
"description": "Repository subfolder (default: auto-detected)",
"type": "string"
}
},
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "git diff of the working folder (unstaged + staged). Optional file to diff a single path.",
"name": "git_diff",
"parameters": {
"properties": {
"dir": {
"description": "Repository subfolder (default: auto-detected)",
"type": "string"
},
"file": {
"description": "Limit the diff to this path",
"type": "string"
}
},
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Clone a git repository (https:// or git@) INTO the working folder of this discussion.",
"name": "git_clone",
"parameters": {
"properties": {
"dir": {
"description": "Target folder name (default: repo name)",
"type": "string"
},
"url": {
"description": "Repository URL",
"type": "string"
}
},
"required": [
"url"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Manage the acceptance criteria of the current task (the contract that defines DONE). action=add posits new criteria (texts[]), action=set updates one (id + status, optional note), action=list shows them, action=clear removes them all. Set criteria BEFORE building. Builder: mark each criterion completed once you have done AND checked it — verification only starts when nothing is left pending. Only the verification pass may mark passed or failed.",
"name": "criteria",
"parameters": {
"properties": {
"action": {
"enum": [
"add",
"set",
"list",
"clear"
],
"type": "string"
},
"id": {
"description": "set: criterion id",
"type": "integer"
},
"note": {
"description": "set: why it failed / how it was verified",
"type": "string"
},
"status": {
"enum": [
"pending",
"completed",
"passed",
"failed"
],
"type": "string"
},
"texts": {
"description": "add: one entry per criterion, short and testable",
"items": {
"type": "string"
},
"type": "array"
}
},
"required": [
"action"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Délègue une recherche ou une relecture à un rôle qui travaille dans son propre contexte et ne rend que sa réponse. Lecture seule.",
"name": "subagent",
"parameters": {
"properties": {
"role": {
"description": "explorer (trouver/cartographier), code-reviewer (relire un changement), planner (découper)",
"enum": [
"explorer",
"code-reviewer",
"planner"
],
"type": "string"
},
"task": {
"description": "La question, précise et autonome : le rôle ne voit pas la discussion.",
"type": "string"
}
},
"required": [
"role",
"task"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Photographie une page web (JS exécuté) pour la MONTRER. La réponse donne la ligne markdown à recopier. Tu ne vois pas l'image.",
"name": "web_screenshot",
"parameters": {
"properties": {
"full_page": {
"description": "Page entière (lourd). Défaut false.",
"type": "boolean"
},
"url": {
"description": "URL complète",
"type": "string"
}
},
"required": [
"url"
],
"type": "object"
}
},
"type": "function"
}
]
},
{
"messages": [
{
"content": "You are Loki, an expert assistant operating directly on this machine with real tools.\n\nYou are in CODE MODE: an autonomous coding agent working in this discussion's folder.\n\nMethod — in this order:\n1. If the task has no acceptance criteria yet, set them FIRST with the criteria tool (action=add): 2-6 short, testable statements of what DONE means. For a trivial task (one obvious change), skip criteria and just do it.\n2. Explore before you change: read the relevant files (read, grep, glob, git_status). Never edit a file you have not read.\n3. Implement with edit (small patches) or write (new files). Match the style of the surrounding code.\n4. Prove it: run the build/tests with bash. A change that was never run is not done. For a web page or UI, start the dev server with bash_bg then use web_screenshot on its localhost URL — a headless Chromium is ALREADY installed, so NEVER install puppeteer, playwright or a browser.\n5. Report briefly what changed and how you verified it.\n\nRules:\n- PLAN ONCE. If the task needs a plan, write it to PLAN.md in ONE short write, then follow it. Between tool calls, think one or two sentences at most — NEVER restate or re-derive the plan: it is in PLAN.md and in your criteria, read them instead.\n- Write each file COMPLETE in a single write call. Many small writes waste turns.\n- When a criterion is done AND checked, mark it completed (criteria set). Verification starts once none is left pending; only it may mark passed.\n- If a command fails, read the error and fix the cause; do not retry the same command unchanged.\n- Ask the user (ask tool) only for decisions that are genuinely theirs; decide the rest yourself.\n- Making the change is your job; shipping it is not. Without an explicit request, never commit, push, reset or rebase, never deploy, restart a service or replace a running binary. Read-only inspection (git_status, git_diff) is fine and encouraged.\n\nThe shell is bash: use its syntax.\nFor anything about the system or files, use bash instead of guessing. Act immediately — call the right tool, then answer. Never end your turn after only thinking. Be concise.\nTo give the user a file, link it in Markdown with its path relative to your working directory — [le rapport](rapport.pdf) — which downloads it. A raw server path is useless: they read you in a browser.\nYou can schedule work for yourself with task_create — only for what must recur or happen later, never a one-off you can do now. A script from your scripts folder runs as a task with no model at all.\n\nDate: {{DATE}}\n\nMachine: host={{HOST}}, cwd={{TMP}}/workspace/discussions/{{CONV}}. This is your working folder: relative paths in write/edit/bash resolve here, and it is the DEFAULT place for scratch work — notes, outputs, a clone, a test. But it is DISPOSABLE: deleting the discussion wipes it. Any script you want to KEEP (or schedule), write it into your scripts folder {{TMP}}/scripts instead — a separate folder a workspace wipe won't touch; you write and run scripts there normally. Do NOT install or write files into system directories such as /usr/local/bin, /usr, /bin or /etc: those need root and are not yours. Only use an absolute path outside this folder (except your scripts folder) when the user explicitly named that location.",
"role": "system"
},
{
"content": "Construis le jeu",
"role": "user"
},
{
"content": "r0",
"role": "assistant"
},
{
"content": "Not done yet — these acceptance criteria are still open:\n[ ] #1 le build passe\n\nContinue working on them. Once one is done and checked, mark it completed (criteria action=set, status=completed). If you need the user's decision, use ask.",
"role": "user"
}
],
"model": "loki",
"parallel_tool_calls": false,
"stream": true,
"stream_options": {
"include_usage": true
},
"temperature": 0.7,
"tools": [
{
"function": {
"description": "Run a shell command (bash syntax) and return stdout, stderr and exit code: inspect the system, read files and logs, run scripts. To CREATE or REWRITE a file use write instead — never echo/cat/python -c. Avoid destructive commands unless asked.",
"name": "bash",
"parameters": {
"properties": {
"command": {
"description": "The command",
"type": "string"
},
"timeout": {
"description": "Timeout s (default 30, max 300)",
"type": "integer"
}
},
"required": [
"command"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Create or replace a file with the exact content given (parent dirs created, content verbatim, no escaping). ALWAYS use this for a script or any text file — NEVER build one through the shell with echo, cat, python -c or Set-Content: quoting breaks.",
"name": "write",
"parameters": {
"properties": {
"content": {
"description": "Full content",
"type": "string"
},
"file": {
"description": "Path",
"type": "string"
}
},
"required": [
"file",
"content"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Patch a file by exact replacement: old → new. old must appear EXACTLY once (add context to make it unique). Prefer this over rewriting a whole file.",
"name": "edit",
"parameters": {
"properties": {
"file": {
"description": "Path",
"type": "string"
},
"new": {
"description": "Replacement",
"type": "string"
},
"old": {
"description": "Exact text to replace (unique)",
"type": "string"
}
},
"required": [
"file",
"old",
"new"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Bring back verbatim an earlier block archived when this conversation was compacted, from the id shown as recall:rN in the summary.",
"name": "recall",
"parameters": {
"properties": {
"id": {
"description": "Block id, e.g. \"r7\".",
"type": "string"
}
},
"required": [
"id"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Find an archived block and its recall id by keywords, when it is not listed in the current summary. Then call recall(id).",
"name": "recall_search",
"parameters": {
"properties": {
"limit": {
"description": "Default 5, max 20",
"type": "integer"
},
"query": {
"description": "Keywords describing the block.",
"type": "string"
}
},
"required": [
"query"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "List your scheduled tasks with their state, next run, and last result. Pass an id to get that task's full last report.",
"name": "task_list",
"parameters": {
"properties": {
"id": {
"description": "Show only this task, with its full last report",
"type": "string"
}
},
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Schedule work to run later, once or repeating. Give 'prompt' for an AI task, OR 'script' (a saved script name) for a script task that runs with no model and no tokens.",
"name": "task_create",
"parameters": {
"properties": {
"at": {
"description": "One-time task at this local time: \"HH:MM\" (next occurrence) or \"YYYY-MM-DD HH:MM\"",
"type": "string"
},
"enabled": {
"description": "Start active (default true)",
"type": "boolean"
},
"in_minutes": {
"description": "One-time task, N minutes from now",
"type": "integer"
},
"name": {
"type": "string"
},
"prompt": {
"description": "AI task: self-contained instruction to run each time (omit if using script)",
"type": "string"
},
"schedule": {
"description": "Repeating: \"@every 2h\" (interval), \"@every 1d@09:00\" (daily time), or a 5-field cron \"min hour dom mon dow\". Omit for a one-time task (use in_minutes or at)",
"type": "string"
},
"script": {
"description": "Script task: name of a script file you put in the scripts folder, run each time with no AI involved",
"type": "string"
}
},
"required": [
"name"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Change one or more fields of a scheduled task by id.",
"name": "task_update",
"parameters": {
"properties": {
"enabled": {
"description": "false pauses, true resumes",
"type": "boolean"
},
"id": {
"type": "string"
},
"name": {
"type": "string"
},
"prompt": {
"type": "string"
},
"schedule": {
"type": "string"
}
},
"required": [
"id"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Delete a scheduled task by id.",
"name": "task_delete",
"parameters": {
"properties": {
"id": {
"type": "string"
}
},
"required": [
"id"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Read a file with 1-indexed line numbers (offset/limit for long files). ALWAYS use this instead of cat/type: output is bounded and reading a file is REQUIRED before you may edit or overwrite it.",
"name": "read",
"parameters": {
"properties": {
"file": {
"description": "Path (relative to the working folder)",
"type": "string"
},
"limit": {
"description": "Lines (default 400, max 1000)",
"type": "integer"
},
"offset": {
"description": "Start line (default 1)",
"type": "integer"
}
},
"required": [
"file"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Search file contents with a Go regular expression. Returns path:line: text matches (bounded). path may be a file or a directory (default: whole working folder).",
"name": "grep",
"parameters": {
"properties": {
"glob": {
"description": "Only files matching this glob (e.g. *.go)",
"type": "string"
},
"path": {
"description": "File or directory to search (default .)",
"type": "string"
},
"pattern": {
"description": "Go regexp",
"type": "string"
}
},
"required": [
"pattern"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "List files matching a glob pattern (** supported, e.g. **/*.go), most recently modified first.",
"name": "glob",
"parameters": {
"properties": {
"pattern": {
"description": "Glob, e.g. src/**/*.ts",
"type": "string"
}
},
"required": [
"pattern"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Ask the user a clarifying question with 2-4 short options, then END your turn. Their answer arrives as the next user message. Use when a decision is genuinely theirs (scope, destructive change, ambiguous requirement) — not for things you can check yourself.",
"name": "ask",
"parameters": {
"properties": {
"options": {
"description": "2-4 short answer options",
"items": {
"type": "string"
},
"type": "array"
},
"question": {
"description": "The question",
"type": "string"
}
},
"required": [
"question"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Start a shell command in the BACKGROUND (dev server, watcher, long build) and return immediately with a job id. Output goes to a log file; check it with bash_tail. Use this instead of appending '&' in bash.",
"name": "bash_bg",
"parameters": {
"properties": {
"command": {
"description": "The command",
"type": "string"
}
},
"required": [
"command"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Read the last lines of a background job's output and its status (running/exited). kill:true stops the job.",
"name": "bash_tail",
"parameters": {
"properties": {
"id": {
"description": "Job id from bash_bg",
"type": "string"
},
"kill": {
"description": "Stop the job",
"type": "boolean"
},
"lines": {
"description": "Lines to show (default 40, max 200)",
"type": "integer"
}
},
"required": [
"id"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "git status of the working folder (branch + changed files, porcelain format).",
"name": "git_status",
"parameters": {
"properties": {
"dir": {
"description": "Repository subfolder (default: auto-detected)",
"type": "string"
}
},
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "git diff of the working folder (unstaged + staged). Optional file to diff a single path.",
"name": "git_diff",
"parameters": {
"properties": {
"dir": {
"description": "Repository subfolder (default: auto-detected)",
"type": "string"
},
"file": {
"description": "Limit the diff to this path",
"type": "string"
}
},
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Clone a git repository (https:// or git@) INTO the working folder of this discussion.",
"name": "git_clone",
"parameters": {
"properties": {
"dir": {
"description": "Target folder name (default: repo name)",
"type": "string"
},
"url": {
"description": "Repository URL",
"type": "string"
}
},
"required": [
"url"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Manage the acceptance criteria of the current task (the contract that defines DONE). action=add posits new criteria (texts[]), action=set updates one (id + status, optional note), action=list shows them, action=clear removes them all. Set criteria BEFORE building. Builder: mark each criterion completed once you have done AND checked it — verification only starts when nothing is left pending. Only the verification pass may mark passed or failed.",
"name": "criteria",
"parameters": {
"properties": {
"action": {
"enum": [
"add",
"set",
"list",
"clear"
],
"type": "string"
},
"id": {
"description": "set: criterion id",
"type": "integer"
},
"note": {
"description": "set: why it failed / how it was verified",
"type": "string"
},
"status": {
"enum": [
"pending",
"completed",
"passed",
"failed"
],
"type": "string"
},
"texts": {
"description": "add: one entry per criterion, short and testable",
"items": {
"type": "string"
},
"type": "array"
}
},
"required": [
"action"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Délègue une recherche ou une relecture à un rôle qui travaille dans son propre contexte et ne rend que sa réponse. Lecture seule.",
"name": "subagent",
"parameters": {
"properties": {
"role": {
"description": "explorer (trouver/cartographier), code-reviewer (relire un changement), planner (découper)",
"enum": [
"explorer",
"code-reviewer",
"planner"
],
"type": "string"
},
"task": {
"description": "La question, précise et autonome : le rôle ne voit pas la discussion.",
"type": "string"
}
},
"required": [
"role",
"task"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Photographie une page web (JS exécuté) pour la MONTRER. La réponse donne la ligne markdown à recopier. Tu ne vois pas l'image.",
"name": "web_screenshot",
"parameters": {
"properties": {
"full_page": {
"description": "Page entière (lourd). Défaut false.",
"type": "boolean"
},
"url": {
"description": "URL complète",
"type": "string"
}
},
"required": [
"url"
],
"type": "object"
}
},
"type": "function"
}
]
},
{
"messages": [
{
"content": "You are Loki, an expert assistant operating directly on this machine with real tools.\n\nYou are in CODE MODE: an autonomous coding agent working in this discussion's folder.\n\nMethod — in this order:\n1. If the task has no acceptance criteria yet, set them FIRST with the criteria tool (action=add): 2-6 short, testable statements of what DONE means. For a trivial task (one obvious change), skip criteria and just do it.\n2. Explore before you change: read the relevant files (read, grep, glob, git_status). Never edit a file you have not read.\n3. Implement with edit (small patches) or write (new files). Match the style of the surrounding code.\n4. Prove it: run the build/tests with bash. A change that was never run is not done. For a web page or UI, start the dev server with bash_bg then use web_screenshot on its localhost URL — a headless Chromium is ALREADY installed, so NEVER install puppeteer, playwright or a browser.\n5. Report briefly what changed and how you verified it.\n\nRules:\n- PLAN ONCE. If the task needs a plan, write it to PLAN.md in ONE short write, then follow it. Between tool calls, think one or two sentences at most — NEVER restate or re-derive the plan: it is in PLAN.md and in your criteria, read them instead.\n- Write each file COMPLETE in a single write call. Many small writes waste turns.\n- When a criterion is done AND checked, mark it completed (criteria set). Verification starts once none is left pending; only it may mark passed.\n- If a command fails, read the error and fix the cause; do not retry the same command unchanged.\n- Ask the user (ask tool) only for decisions that are genuinely theirs; decide the rest yourself.\n- Making the change is your job; shipping it is not. Without an explicit request, never commit, push, reset or rebase, never deploy, restart a service or replace a running binary. Read-only inspection (git_status, git_diff) is fine and encouraged.\n\nThe shell is bash: use its syntax.\nFor anything about the system or files, use bash instead of guessing. Act immediately — call the right tool, then answer. Never end your turn after only thinking. Be concise.\nTo give the user a file, link it in Markdown with its path relative to your working directory — [le rapport](rapport.pdf) — which downloads it. A raw server path is useless: they read you in a browser.\nYou can schedule work for yourself with task_create — only for what must recur or happen later, never a one-off you can do now. A script from your scripts folder runs as a task with no model at all.\n\nDate: {{DATE}}\n\nMachine: host={{HOST}}, cwd={{TMP}}/workspace/discussions/{{CONV}}. This is your working folder: relative paths in write/edit/bash resolve here, and it is the DEFAULT place for scratch work — notes, outputs, a clone, a test. But it is DISPOSABLE: deleting the discussion wipes it. Any script you want to KEEP (or schedule), write it into your scripts folder {{TMP}}/scripts instead — a separate folder a workspace wipe won't touch; you write and run scripts there normally. Do NOT install or write files into system directories such as /usr/local/bin, /usr, /bin or /etc: those need root and are not yours. Only use an absolute path outside this folder (except your scripts folder) when the user explicitly named that location.",
"role": "system"
},
{
"content": "Construis le jeu",
"role": "user"
},
{
"content": "r0",
"role": "assistant"
},
{
"content": "Not done yet — these acceptance criteria are still open:\n[ ] #1 le build passe\n\nContinue working on them. Once one is done and checked, mark it completed (criteria action=set, status=completed). If you need the user's decision, use ask.",
"role": "user"
},
{
"content": "r1",
"role": "assistant"
},
{
"content": "Not done yet — these acceptance criteria are still open:\n[ ] #1 le build passe\n\nContinue working on them. Once one is done and checked, mark it completed (criteria action=set, status=completed). If you need the user's decision, use ask.",
"role": "user"
}
],
"model": "loki",
"parallel_tool_calls": false,
"stream": true,
"stream_options": {
"include_usage": true
},
"temperature": 0.7,
"tools": [
{
"function": {
"description": "Run a shell command (bash syntax) and return stdout, stderr and exit code: inspect the system, read files and logs, run scripts. To CREATE or REWRITE a file use write instead — never echo/cat/python -c. Avoid destructive commands unless asked.",
"name": "bash",
"parameters": {
"properties": {
"command": {
"description": "The command",
"type": "string"
},
"timeout": {
"description": "Timeout s (default 30, max 300)",
"type": "integer"
}
},
"required": [
"command"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Create or replace a file with the exact content given (parent dirs created, content verbatim, no escaping). ALWAYS use this for a script or any text file — NEVER build one through the shell with echo, cat, python -c or Set-Content: quoting breaks.",
"name": "write",
"parameters": {
"properties": {
"content": {
"description": "Full content",
"type": "string"
},
"file": {
"description": "Path",
"type": "string"
}
},
"required": [
"file",
"content"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Patch a file by exact replacement: old → new. old must appear EXACTLY once (add context to make it unique). Prefer this over rewriting a whole file.",
"name": "edit",
"parameters": {
"properties": {
"file": {
"description": "Path",
"type": "string"
},
"new": {
"description": "Replacement",
"type": "string"
},
"old": {
"description": "Exact text to replace (unique)",
"type": "string"
}
},
"required": [
"file",
"old",
"new"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Bring back verbatim an earlier block archived when this conversation was compacted, from the id shown as recall:rN in the summary.",
"name": "recall",
"parameters": {
"properties": {
"id": {
"description": "Block id, e.g. \"r7\".",
"type": "string"
}
},
"required": [
"id"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Find an archived block and its recall id by keywords, when it is not listed in the current summary. Then call recall(id).",
"name": "recall_search",
"parameters": {
"properties": {
"limit": {
"description": "Default 5, max 20",
"type": "integer"
},
"query": {
"description": "Keywords describing the block.",
"type": "string"
}
},
"required": [
"query"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "List your scheduled tasks with their state, next run, and last result. Pass an id to get that task's full last report.",
"name": "task_list",
"parameters": {
"properties": {
"id": {
"description": "Show only this task, with its full last report",
"type": "string"
}
},
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Schedule work to run later, once or repeating. Give 'prompt' for an AI task, OR 'script' (a saved script name) for a script task that runs with no model and no tokens.",
"name": "task_create",
"parameters": {
"properties": {
"at": {
"description": "One-time task at this local time: \"HH:MM\" (next occurrence) or \"YYYY-MM-DD HH:MM\"",
"type": "string"
},
"enabled": {
"description": "Start active (default true)",
"type": "boolean"
},
"in_minutes": {
"description": "One-time task, N minutes from now",
"type": "integer"
},
"name": {
"type": "string"
},
"prompt": {
"description": "AI task: self-contained instruction to run each time (omit if using script)",
"type": "string"
},
"schedule": {
"description": "Repeating: \"@every 2h\" (interval), \"@every 1d@09:00\" (daily time), or a 5-field cron \"min hour dom mon dow\". Omit for a one-time task (use in_minutes or at)",
"type": "string"
},
"script": {
"description": "Script task: name of a script file you put in the scripts folder, run each time with no AI involved",
"type": "string"
}
},
"required": [
"name"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Change one or more fields of a scheduled task by id.",
"name": "task_update",
"parameters": {
"properties": {
"enabled": {
"description": "false pauses, true resumes",
"type": "boolean"
},
"id": {
"type": "string"
},
"name": {
"type": "string"
},
"prompt": {
"type": "string"
},
"schedule": {
"type": "string"
}
},
"required": [
"id"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Delete a scheduled task by id.",
"name": "task_delete",
"parameters": {
"properties": {
"id": {
"type": "string"
}
},
"required": [
"id"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Read a file with 1-indexed line numbers (offset/limit for long files). ALWAYS use this instead of cat/type: output is bounded and reading a file is REQUIRED before you may edit or overwrite it.",
"name": "read",
"parameters": {
"properties": {
"file": {
"description": "Path (relative to the working folder)",
"type": "string"
},
"limit": {
"description": "Lines (default 400, max 1000)",
"type": "integer"
},
"offset": {
"description": "Start line (default 1)",
"type": "integer"
}
},
"required": [
"file"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Search file contents with a Go regular expression. Returns path:line: text matches (bounded). path may be a file or a directory (default: whole working folder).",
"name": "grep",
"parameters": {
"properties": {
"glob": {
"description": "Only files matching this glob (e.g. *.go)",
"type": "string"
},
"path": {
"description": "File or directory to search (default .)",
"type": "string"
},
"pattern": {
"description": "Go regexp",
"type": "string"
}
},
"required": [
"pattern"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "List files matching a glob pattern (** supported, e.g. **/*.go), most recently modified first.",
"name": "glob",
"parameters": {
"properties": {
"pattern": {
"description": "Glob, e.g. src/**/*.ts",
"type": "string"
}
},
"required": [
"pattern"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Ask the user a clarifying question with 2-4 short options, then END your turn. Their answer arrives as the next user message. Use when a decision is genuinely theirs (scope, destructive change, ambiguous requirement) — not for things you can check yourself.",
"name": "ask",
"parameters": {
"properties": {
"options": {
"description": "2-4 short answer options",
"items": {
"type": "string"
},
"type": "array"
},
"question": {
"description": "The question",
"type": "string"
}
},
"required": [
"question"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Start a shell command in the BACKGROUND (dev server, watcher, long build) and return immediately with a job id. Output goes to a log file; check it with bash_tail. Use this instead of appending '&' in bash.",
"name": "bash_bg",
"parameters": {
"properties": {
"command": {
"description": "The command",
"type": "string"
}
},
"required": [
"command"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Read the last lines of a background job's output and its status (running/exited). kill:true stops the job.",
"name": "bash_tail",
"parameters": {
"properties": {
"id": {
"description": "Job id from bash_bg",
"type": "string"
},
"kill": {
"description": "Stop the job",
"type": "boolean"
},
"lines": {
"description": "Lines to show (default 40, max 200)",
"type": "integer"
}
},
"required": [
"id"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "git status of the working folder (branch + changed files, porcelain format).",
"name": "git_status",
"parameters": {
"properties": {
"dir": {
"description": "Repository subfolder (default: auto-detected)",
"type": "string"
}
},
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "git diff of the working folder (unstaged + staged). Optional file to diff a single path.",
"name": "git_diff",
"parameters": {
"properties": {
"dir": {
"description": "Repository subfolder (default: auto-detected)",
"type": "string"
},
"file": {
"description": "Limit the diff to this path",
"type": "string"
}
},
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Clone a git repository (https:// or git@) INTO the working folder of this discussion.",
"name": "git_clone",
"parameters": {
"properties": {
"dir": {
"description": "Target folder name (default: repo name)",
"type": "string"
},
"url": {
"description": "Repository URL",
"type": "string"
}
},
"required": [
"url"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Manage the acceptance criteria of the current task (the contract that defines DONE). action=add posits new criteria (texts[]), action=set updates one (id + status, optional note), action=list shows them, action=clear removes them all. Set criteria BEFORE building. Builder: mark each criterion completed once you have done AND checked it — verification only starts when nothing is left pending. Only the verification pass may mark passed or failed.",
"name": "criteria",
"parameters": {
"properties": {
"action": {
"enum": [
"add",
"set",
"list",
"clear"
],
"type": "string"
},
"id": {
"description": "set: criterion id",
"type": "integer"
},
"note": {
"description": "set: why it failed / how it was verified",
"type": "string"
},
"status": {
"enum": [
"pending",
"completed",
"passed",
"failed"
],
"type": "string"
},
"texts": {
"description": "add: one entry per criterion, short and testable",
"items": {
"type": "string"
},
"type": "array"
}
},
"required": [
"action"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Délègue une recherche ou une relecture à un rôle qui travaille dans son propre contexte et ne rend que sa réponse. Lecture seule.",
"name": "subagent",
"parameters": {
"properties": {
"role": {
"description": "explorer (trouver/cartographier), code-reviewer (relire un changement), planner (découper)",
"enum": [
"explorer",
"code-reviewer",
"planner"
],
"type": "string"
},
"task": {
"description": "La question, précise et autonome : le rôle ne voit pas la discussion.",
"type": "string"
}
},
"required": [
"role",
"task"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Photographie une page web (JS exécuté) pour la MONTRER. La réponse donne la ligne markdown à recopier. Tu ne vois pas l'image.",
"name": "web_screenshot",
"parameters": {
"properties": {
"full_page": {
"description": "Page entière (lourd). Défaut false.",
"type": "boolean"
},
"url": {
"description": "URL complète",
"type": "string"
}
},
"required": [
"url"
],
"type": "object"
}
},
"type": "function"
}
]
},
{
"messages": [
{
"content": "You are the VERIFICATION pass of code mode. You independently check the acceptance criteria against the actual code — you did not write it, trust nothing.\n\nFor each criterion still pending or failed:\n1. Understand what it actually requires.\n2. Check the real changes: git_diff, read the touched files, run the relevant build/test command with bash when applicable.\n3. Record the verdict with the criteria tool: action=set, status=passed (satisfied) or failed (with a note saying exactly what is missing — actionable, one line).\n\nRules:\n- Trivial or non-code criteria that are plainly satisfied: pass them without exploring.\n- Only fail a criterion that genuinely misses its requirement; a different-but-valid implementation passes.\n- Do NOT fix anything yourself. You only verify and report.\n- Be fast: check the diff first, read only what the criteria require.\n\nMachine: host={{HOST}}, cwd={{TMP}}/workspace/discussions/{{CONV}}. This is your working folder: relative paths in write/edit/bash resolve here, and it is the DEFAULT place for scratch work — notes, outputs, a clone, a test. But it is DISPOSABLE: deleting the discussion wipes it. Any script you want to KEEP (or schedule), write it into your scripts folder {{TMP}}/scripts instead — a separate folder a workspace wipe won't touch; you write and run scripts there normally. Do NOT install or write files into system directories such as /usr/local/bin, /usr, /bin or /etc: those need root and are not yours. Only use an absolute path outside this folder (except your scripts folder) when the user explicitly named that location.",
"role": "system"
},
{
"content": "Task:\nNot done yet — these acceptance criteria are still open:\n[ ] #1 le build passe\n\nContinue working on them. Once one is done and checked, mark it completed (criteria action=set, status=completed). If you need the user's decision, use ask.\n\nAcceptance criteria:\n[ ] #1 le build passe\n\nVerify each non-passed criterion now (git_diff first), and record every verdict with the criteria tool.",
"role": "user"
}
],
"model": "loki",
"parallel_tool_calls": false,
"stream": true,
"stream_options": {
"include_usage": true
},
"temperature": 0.7,
"tools": [
{
"function": {
"description": "Run a shell command (bash syntax) and return stdout, stderr and exit code: inspect the system, read files and logs, run scripts. To CREATE or REWRITE a file use write instead — never echo/cat/python -c. Avoid destructive commands unless asked.",
"name": "bash",
"parameters": {
"properties": {
"command": {
"description": "The command",
"type": "string"
},
"timeout": {
"description": "Timeout s (default 30, max 300)",
"type": "integer"
}
},
"required": [
"command"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Read a file with 1-indexed line numbers (offset/limit for long files). ALWAYS use this instead of cat/type: output is bounded and reading a file is REQUIRED before you may edit or overwrite it.",
"name": "read",
"parameters": {
"properties": {
"file": {
"description": "Path (relative to the working folder)",
"type": "string"
},
"limit": {
"description": "Lines (default 400, max 1000)",
"type": "integer"
},
"offset": {
"description": "Start line (default 1)",
"type": "integer"
}
},
"required": [
"file"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Search file contents with a Go regular expression. Returns path:line: text matches (bounded). path may be a file or a directory (default: whole working folder).",
"name": "grep",
"parameters": {
"properties": {
"glob": {
"description": "Only files matching this glob (e.g. *.go)",
"type": "string"
},
"path": {
"description": "File or directory to search (default .)",
"type": "string"
},
"pattern": {
"description": "Go regexp",
"type": "string"
}
},
"required": [
"pattern"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "List files matching a glob pattern (** supported, e.g. **/*.go), most recently modified first.",
"name": "glob",
"parameters": {
"properties": {
"pattern": {
"description": "Glob, e.g. src/**/*.ts",
"type": "string"
}
},
"required": [
"pattern"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "git status of the working folder (branch + changed files, porcelain format).",
"name": "git_status",
"parameters": {
"properties": {
"dir": {
"description": "Repository subfolder (default: auto-detected)",
"type": "string"
}
},
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "git diff of the working folder (unstaged + staged). Optional file to diff a single path.",
"name": "git_diff",
"parameters": {
"properties": {
"dir": {
"description": "Repository subfolder (default: auto-detected)",
"type": "string"
},
"file": {
"description": "Limit the diff to this path",
"type": "string"
}
},
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Manage the acceptance criteria of the current task (the contract that defines DONE). action=add posits new criteria (texts[]), action=set updates one (id + status, optional note), action=list shows them, action=clear removes them all. Set criteria BEFORE building. Builder: mark each criterion completed once you have done AND checked it — verification only starts when nothing is left pending. Only the verification pass may mark passed or failed.",
"name": "criteria",
"parameters": {
"properties": {
"action": {
"enum": [
"add",
"set",
"list",
"clear"
],
"type": "string"
},
"id": {
"description": "set: criterion id",
"type": "integer"
},
"note": {
"description": "set: why it failed / how it was verified",
"type": "string"
},
"status": {
"enum": [
"pending",
"completed",
"passed",
"failed"
],
"type": "string"
},
"texts": {
"description": "add: one entry per criterion, short and testable",
"items": {
"type": "string"
},
"type": "array"
}
},
"required": [
"action"
],
"type": "object"
}
},
"type": "function"
}
]
},
{
"messages": [
{
"content": "You are Loki, an expert assistant operating directly on this machine with real tools.\n\nYou are in CODE MODE: an autonomous coding agent working in this discussion's folder.\n\nMethod — in this order:\n1. If the task has no acceptance criteria yet, set them FIRST with the criteria tool (action=add): 2-6 short, testable statements of what DONE means. For a trivial task (one obvious change), skip criteria and just do it.\n2. Explore before you change: read the relevant files (read, grep, glob, git_status). Never edit a file you have not read.\n3. Implement with edit (small patches) or write (new files). Match the style of the surrounding code.\n4. Prove it: run the build/tests with bash. A change that was never run is not done. For a web page or UI, start the dev server with bash_bg then use web_screenshot on its localhost URL — a headless Chromium is ALREADY installed, so NEVER install puppeteer, playwright or a browser.\n5. Report briefly what changed and how you verified it.\n\nRules:\n- PLAN ONCE. If the task needs a plan, write it to PLAN.md in ONE short write, then follow it. Between tool calls, think one or two sentences at most — NEVER restate or re-derive the plan: it is in PLAN.md and in your criteria, read them instead.\n- Write each file COMPLETE in a single write call. Many small writes waste turns.\n- When a criterion is done AND checked, mark it completed (criteria set). Verification starts once none is left pending; only it may mark passed.\n- If a command fails, read the error and fix the cause; do not retry the same command unchanged.\n- Ask the user (ask tool) only for decisions that are genuinely theirs; decide the rest yourself.\n- Making the change is your job; shipping it is not. Without an explicit request, never commit, push, reset or rebase, never deploy, restart a service or replace a running binary. Read-only inspection (git_status, git_diff) is fine and encouraged.\n\nThe shell is bash: use its syntax.\nFor anything about the system or files, use bash instead of guessing. Act immediately — call the right tool, then answer. Never end your turn after only thinking. Be concise.\nTo give the user a file, link it in Markdown with its path relative to your working directory — [le rapport](rapport.pdf) — which downloads it. A raw server path is useless: they read you in a browser.\nYou can schedule work for yourself with task_create — only for what must recur or happen later, never a one-off you can do now. A script from your scripts folder runs as a task with no model at all.\n\nDate: {{DATE}}\n\nMachine: host={{HOST}}, cwd={{TMP}}/workspace/discussions/{{CONV}}. This is your working folder: relative paths in write/edit/bash resolve here, and it is the DEFAULT place for scratch work — notes, outputs, a clone, a test. But it is DISPOSABLE: deleting the discussion wipes it. Any script you want to KEEP (or schedule), write it into your scripts folder {{TMP}}/scripts instead — a separate folder a workspace wipe won't touch; you write and run scripts there normally. Do NOT install or write files into system directories such as /usr/local/bin, /usr, /bin or /etc: those need root and are not yours. Only use an absolute path outside this folder (except your scripts folder) when the user explicitly named that location.",
"role": "system"
},
{
"content": "Construis le jeu",
"role": "user"
},
{
"content": "r0",
"role": "assistant"
},
{
"content": "Not done yet — these acceptance criteria are still open:\n[ ] #1 le build passe\n\nContinue working on them. Once one is done and checked, mark it completed (criteria action=set, status=completed). If you need the user's decision, use ask.",
"role": "user"
},
{
"content": "r1",
"role": "assistant"
},
{
"content": "Not done yet — these acceptance criteria are still open:\n[ ] #1 le build passe\n\nContinue working on them. Once one is done and checked, mark it completed (criteria action=set, status=completed). If you need the user's decision, use ask.",
"role": "user"
},
{
"content": "r2",
"role": "assistant"
},
{
"content": "Verification failed on these criteria:\n[ ] #1 le build passe\n\nFix the code so they pass. Address the notes precisely; do not touch what already passed. Mark each fixed criterion completed again.",
"role": "user"
}
],
"model": "loki",
"parallel_tool_calls": false,
"stream": true,
"stream_options": {
"include_usage": true
},
"temperature": 0.7,
"tools": [
{
"function": {
"description": "Run a shell command (bash syntax) and return stdout, stderr and exit code: inspect the system, read files and logs, run scripts. To CREATE or REWRITE a file use write instead — never echo/cat/python -c. Avoid destructive commands unless asked.",
"name": "bash",
"parameters": {
"properties": {
"command": {
"description": "The command",
"type": "string"
},
"timeout": {
"description": "Timeout s (default 30, max 300)",
"type": "integer"
}
},
"required": [
"command"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Create or replace a file with the exact content given (parent dirs created, content verbatim, no escaping). ALWAYS use this for a script or any text file — NEVER build one through the shell with echo, cat, python -c or Set-Content: quoting breaks.",
"name": "write",
"parameters": {
"properties": {
"content": {
"description": "Full content",
"type": "string"
},
"file": {
"description": "Path",
"type": "string"
}
},
"required": [
"file",
"content"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Patch a file by exact replacement: old → new. old must appear EXACTLY once (add context to make it unique). Prefer this over rewriting a whole file.",
"name": "edit",
"parameters": {
"properties": {
"file": {
"description": "Path",
"type": "string"
},
"new": {
"description": "Replacement",
"type": "string"
},
"old": {
"description": "Exact text to replace (unique)",
"type": "string"
}
},
"required": [
"file",
"old",
"new"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Bring back verbatim an earlier block archived when this conversation was compacted, from the id shown as recall:rN in the summary.",
"name": "recall",
"parameters": {
"properties": {
"id": {
"description": "Block id, e.g. \"r7\".",
"type": "string"
}
},
"required": [
"id"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Find an archived block and its recall id by keywords, when it is not listed in the current summary. Then call recall(id).",
"name": "recall_search",
"parameters": {
"properties": {
"limit": {
"description": "Default 5, max 20",
"type": "integer"
},
"query": {
"description": "Keywords describing the block.",
"type": "string"
}
},
"required": [
"query"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "List your scheduled tasks with their state, next run, and last result. Pass an id to get that task's full last report.",
"name": "task_list",
"parameters": {
"properties": {
"id": {
"description": "Show only this task, with its full last report",
"type": "string"
}
},
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Schedule work to run later, once or repeating. Give 'prompt' for an AI task, OR 'script' (a saved script name) for a script task that runs with no model and no tokens.",
"name": "task_create",
"parameters": {
"properties": {
"at": {
"description": "One-time task at this local time: \"HH:MM\" (next occurrence) or \"YYYY-MM-DD HH:MM\"",
"type": "string"
},
"enabled": {
"description": "Start active (default true)",
"type": "boolean"
},
"in_minutes": {
"description": "One-time task, N minutes from now",
"type": "integer"
},
"name": {
"type": "string"
},
"prompt": {
"description": "AI task: self-contained instruction to run each time (omit if using script)",
"type": "string"
},
"schedule": {
"description": "Repeating: \"@every 2h\" (interval), \"@every 1d@09:00\" (daily time), or a 5-field cron \"min hour dom mon dow\". Omit for a one-time task (use in_minutes or at)",
"type": "string"
},
"script": {
"description": "Script task: name of a script file you put in the scripts folder, run each time with no AI involved",
"type": "string"
}
},
"required": [
"name"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Change one or more fields of a scheduled task by id.",
"name": "task_update",
"parameters": {
"properties": {
"enabled": {
"description": "false pauses, true resumes",
"type": "boolean"
},
"id": {
"type": "string"
},
"name": {
"type": "string"
},
"prompt": {
"type": "string"
},
"schedule": {
"type": "string"
}
},
"required": [
"id"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Delete a scheduled task by id.",
"name": "task_delete",
"parameters": {
"properties": {
"id": {
"type": "string"
}
},
"required": [
"id"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Read a file with 1-indexed line numbers (offset/limit for long files). ALWAYS use this instead of cat/type: output is bounded and reading a file is REQUIRED before you may edit or overwrite it.",
"name": "read",
"parameters": {
"properties": {
"file": {
"description": "Path (relative to the working folder)",
"type": "string"
},
"limit": {
"description": "Lines (default 400, max 1000)",
"type": "integer"
},
"offset": {
"description": "Start line (default 1)",
"type": "integer"
}
},
"required": [
"file"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Search file contents with a Go regular expression. Returns path:line: text matches (bounded). path may be a file or a directory (default: whole working folder).",
"name": "grep",
"parameters": {
"properties": {
"glob": {
"description": "Only files matching this glob (e.g. *.go)",
"type": "string"
},
"path": {
"description": "File or directory to search (default .)",
"type": "string"
},
"pattern": {
"description": "Go regexp",
"type": "string"
}
},
"required": [
"pattern"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "List files matching a glob pattern (** supported, e.g. **/*.go), most recently modified first.",
"name": "glob",
"parameters": {
"properties": {
"pattern": {
"description": "Glob, e.g. src/**/*.ts",
"type": "string"
}
},
"required": [
"pattern"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Ask the user a clarifying question with 2-4 short options, then END your turn. Their answer arrives as the next user message. Use when a decision is genuinely theirs (scope, destructive change, ambiguous requirement) — not for things you can check yourself.",
"name": "ask",
"parameters": {
"properties": {
"options": {
"description": "2-4 short answer options",
"items": {
"type": "string"
},
"type": "array"
},
"question": {
"description": "The question",
"type": "string"
}
},
"required": [
"question"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Start a shell command in the BACKGROUND (dev server, watcher, long build) and return immediately with a job id. Output goes to a log file; check it with bash_tail. Use this instead of appending '&' in bash.",
"name": "bash_bg",
"parameters": {
"properties": {
"command": {
"description": "The command",
"type": "string"
}
},
"required": [
"command"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Read the last lines of a background job's output and its status (running/exited). kill:true stops the job.",
"name": "bash_tail",
"parameters": {
"properties": {
"id": {
"description": "Job id from bash_bg",
"type": "string"
},
"kill": {
"description": "Stop the job",
"type": "boolean"
},
"lines": {
"description": "Lines to show (default 40, max 200)",
"type": "integer"
}
},
"required": [
"id"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "git status of the working folder (branch + changed files, porcelain format).",
"name": "git_status",
"parameters": {
"properties": {
"dir": {
"description": "Repository subfolder (default: auto-detected)",
"type": "string"
}
},
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "git diff of the working folder (unstaged + staged). Optional file to diff a single path.",
"name": "git_diff",
"parameters": {
"properties": {
"dir": {
"description": "Repository subfolder (default: auto-detected)",
"type": "string"
},
"file": {
"description": "Limit the diff to this path",
"type": "string"
}
},
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Clone a git repository (https:// or git@) INTO the working folder of this discussion.",
"name": "git_clone",
"parameters": {
"properties": {
"dir": {
"description": "Target folder name (default: repo name)",
"type": "string"
},
"url": {
"description": "Repository URL",
"type": "string"
}
},
"required": [
"url"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Manage the acceptance criteria of the current task (the contract that defines DONE). action=add posits new criteria (texts[]), action=set updates one (id + status, optional note), action=list shows them, action=clear removes them all. Set criteria BEFORE building. Builder: mark each criterion completed once you have done AND checked it — verification only starts when nothing is left pending. Only the verification pass may mark passed or failed.",
"name": "criteria",
"parameters": {
"properties": {
"action": {
"enum": [
"add",
"set",
"list",
"clear"
],
"type": "string"
},
"id": {
"description": "set: criterion id",
"type": "integer"
},
"note": {
"description": "set: why it failed / how it was verified",
"type": "string"
},
"status": {
"enum": [
"pending",
"completed",
"passed",
"failed"
],
"type": "string"
},
"texts": {
"description": "add: one entry per criterion, short and testable",
"items": {
"type": "string"
},
"type": "array"
}
},
"required": [
"action"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Délègue une recherche ou une relecture à un rôle qui travaille dans son propre contexte et ne rend que sa réponse. Lecture seule.",
"name": "subagent",
"parameters": {
"properties": {
"role": {
"description": "explorer (trouver/cartographier), code-reviewer (relire un changement), planner (découper)",
"enum": [
"explorer",
"code-reviewer",
"planner"
],
"type": "string"
},
"task": {
"description": "La question, précise et autonome : le rôle ne voit pas la discussion.",
"type": "string"
}
},
"required": [
"role",
"task"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Photographie une page web (JS exécuté) pour la MONTRER. La réponse donne la ligne markdown à recopier. Tu ne vois pas l'image.",
"name": "web_screenshot",
"parameters": {
"properties": {
"full_page": {
"description": "Page entière (lourd). Défaut false.",
"type": "boolean"
},
"url": {
"description": "URL complète",
"type": "string"
}
},
"required": [
"url"
],
"type": "object"
}
},
"type": "function"
}
]
},
{
"messages": [
{
"content": "You are the VERIFICATION pass of code mode. You independently check the acceptance criteria against the actual code — you did not write it, trust nothing.\n\nFor each criterion still pending or failed:\n1. Understand what it actually requires.\n2. Check the real changes: git_diff, read the touched files, run the relevant build/test command with bash when applicable.\n3. Record the verdict with the criteria tool: action=set, status=passed (satisfied) or failed (with a note saying exactly what is missing — actionable, one line).\n\nRules:\n- Trivial or non-code criteria that are plainly satisfied: pass them without exploring.\n- Only fail a criterion that genuinely misses its requirement; a different-but-valid implementation passes.\n- Do NOT fix anything yourself. You only verify and report.\n- Be fast: check the diff first, read only what the criteria require.\n\nMachine: host={{HOST}}, cwd={{TMP}}/workspace/discussions/{{CONV}}. This is your working folder: relative paths in write/edit/bash resolve here, and it is the DEFAULT place for scratch work — notes, outputs, a clone, a test. But it is DISPOSABLE: deleting the discussion wipes it. Any script you want to KEEP (or schedule), write it into your scripts folder {{TMP}}/scripts instead — a separate folder a workspace wipe won't touch; you write and run scripts there normally. Do NOT install or write files into system directories such as /usr/local/bin, /usr, /bin or /etc: those need root and are not yours. Only use an absolute path outside this folder (except your scripts folder) when the user explicitly named that location.",
"role": "system"
},
{
"content": "Task:\nVerification failed on these criteria:\n[ ] #1 le build passe\n\nFix the code so they pass. Address the notes precisely; do not touch what already passed. Mark each fixed criterion completed again.\n\nAcceptance criteria:\n[ ] #1 le build passe\n\nVerify each non-passed criterion now (git_diff first), and record every verdict with the criteria tool.",
"role": "user"
}
],
"model": "loki",
"parallel_tool_calls": false,
"stream": true,
"stream_options": {
"include_usage": true
},
"temperature": 0.7,
"tools": [
{
"function": {
"description": "Run a shell command (bash syntax) and return stdout, stderr and exit code: inspect the system, read files and logs, run scripts. To CREATE or REWRITE a file use write instead — never echo/cat/python -c. Avoid destructive commands unless asked.",
"name": "bash",
"parameters": {
"properties": {
"command": {
"description": "The command",
"type": "string"
},
"timeout": {
"description": "Timeout s (default 30, max 300)",
"type": "integer"
}
},
"required": [
"command"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Read a file with 1-indexed line numbers (offset/limit for long files). ALWAYS use this instead of cat/type: output is bounded and reading a file is REQUIRED before you may edit or overwrite it.",
"name": "read",
"parameters": {
"properties": {
"file": {
"description": "Path (relative to the working folder)",
"type": "string"
},
"limit": {
"description": "Lines (default 400, max 1000)",
"type": "integer"
},
"offset": {
"description": "Start line (default 1)",
"type": "integer"
}
},
"required": [
"file"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Search file contents with a Go regular expression. Returns path:line: text matches (bounded). path may be a file or a directory (default: whole working folder).",
"name": "grep",
"parameters": {
"properties": {
"glob": {
"description": "Only files matching this glob (e.g. *.go)",
"type": "string"
},
"path": {
"description": "File or directory to search (default .)",
"type": "string"
},
"pattern": {
"description": "Go regexp",
"type": "string"
}
},
"required": [
"pattern"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "List files matching a glob pattern (** supported, e.g. **/*.go), most recently modified first.",
"name": "glob",
"parameters": {
"properties": {
"pattern": {
"description": "Glob, e.g. src/**/*.ts",
"type": "string"
}
},
"required": [
"pattern"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "git status of the working folder (branch + changed files, porcelain format).",
"name": "git_status",
"parameters": {
"properties": {
"dir": {
"description": "Repository subfolder (default: auto-detected)",
"type": "string"
}
},
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "git diff of the working folder (unstaged + staged). Optional file to diff a single path.",
"name": "git_diff",
"parameters": {
"properties": {
"dir": {
"description": "Repository subfolder (default: auto-detected)",
"type": "string"
},
"file": {
"description": "Limit the diff to this path",
"type": "string"
}
},
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Manage the acceptance criteria of the current task (the contract that defines DONE). action=add posits new criteria (texts[]), action=set updates one (id + status, optional note), action=list shows them, action=clear removes them all. Set criteria BEFORE building. Builder: mark each criterion completed once you have done AND checked it — verification only starts when nothing is left pending. Only the verification pass may mark passed or failed.",
"name": "criteria",
"parameters": {
"properties": {
"action": {
"enum": [
"add",
"set",
"list",
"clear"
],
"type": "string"
},
"id": {
"description": "set: criterion id",
"type": "integer"
},
"note": {
"description": "set: why it failed / how it was verified",
"type": "string"
},
"status": {
"enum": [
"pending",
"completed",
"passed",
"failed"
],
"type": "string"
},
"texts": {
"description": "add: one entry per criterion, short and testable",
"items": {
"type": "string"
},
"type": "array"
}
},
"required": [
"action"
],
"type": "object"
}
},
"type": "function"
}
]
},
{
"messages": [
{
"content": "You are Loki, an expert assistant operating directly on this machine with real tools.\n\nYou are in CODE MODE: an autonomous coding agent working in this discussion's folder.\n\nMethod — in this order:\n1. If the task has no acceptance criteria yet, set them FIRST with the criteria tool (action=add): 2-6 short, testable statements of what DONE means. For a trivial task (one obvious change), skip criteria and just do it.\n2. Explore before you change: read the relevant files (read, grep, glob, git_status). Never edit a file you have not read.\n3. Implement with edit (small patches) or write (new files). Match the style of the surrounding code.\n4. Prove it: run the build/tests with bash. A change that was never run is not done. For a web page or UI, start the dev server with bash_bg then use web_screenshot on its localhost URL — a headless Chromium is ALREADY installed, so NEVER install puppeteer, playwright or a browser.\n5. Report briefly what changed and how you verified it.\n\nRules:\n- PLAN ONCE. If the task needs a plan, write it to PLAN.md in ONE short write, then follow it. Between tool calls, think one or two sentences at most — NEVER restate or re-derive the plan: it is in PLAN.md and in your criteria, read them instead.\n- Write each file COMPLETE in a single write call. Many small writes waste turns.\n- When a criterion is done AND checked, mark it completed (criteria set). Verification starts once none is left pending; only it may mark passed.\n- If a command fails, read the error and fix the cause; do not retry the same command unchanged.\n- Ask the user (ask tool) only for decisions that are genuinely theirs; decide the rest yourself.\n- Making the change is your job; shipping it is not. Without an explicit request, never commit, push, reset or rebase, never deploy, restart a service or replace a running binary. Read-only inspection (git_status, git_diff) is fine and encouraged.\n\nThe shell is bash: use its syntax.\nFor anything about the system or files, use bash instead of guessing. Act immediately — call the right tool, then answer. Never end your turn after only thinking. Be concise.\nTo give the user a file, link it in Markdown with its path relative to your working directory — [le rapport](rapport.pdf) — which downloads it. A raw server path is useless: they read you in a browser.\nYou can schedule work for yourself with task_create — only for what must recur or happen later, never a one-off you can do now. A script from your scripts folder runs as a task with no model at all.\n\nDate: {{DATE}}\n\nMachine: host={{HOST}}, cwd={{TMP}}/workspace/discussions/{{CONV}}. This is your working folder: relative paths in write/edit/bash resolve here, and it is the DEFAULT place for scratch work — notes, outputs, a clone, a test. But it is DISPOSABLE: deleting the discussion wipes it. Any script you want to KEEP (or schedule), write it into your scripts folder {{TMP}}/scripts instead — a separate folder a workspace wipe won't touch; you write and run scripts there normally. Do NOT install or write files into system directories such as /usr/local/bin, /usr, /bin or /etc: those need root and are not yours. Only use an absolute path outside this folder (except your scripts folder) when the user explicitly named that location.",
"role": "system"
},
{
"content": "Construis le jeu",
"role": "user"
},
{
"content": "r0",
"role": "assistant"
},
{
"content": "Not done yet — these acceptance criteria are still open:\n[ ] #1 le build passe\n\nContinue working on them. Once one is done and checked, mark it completed (criteria action=set, status=completed). If you need the user's decision, use ask.",
"role": "user"
},
{
"content": "r1",
"role": "assistant"
},
{
"content": "Not done yet — these acceptance criteria are still open:\n[ ] #1 le build passe\n\nContinue working on them. Once one is done and checked, mark it completed (criteria action=set, status=completed). If you need the user's decision, use ask.",
"role": "user"
},
{
"content": "r2",
"role": "assistant"
},
{
"content": "Verification failed on these criteria:\n[ ] #1 le build passe\n\nFix the code so they pass. Address the notes precisely; do not touch what already passed. Mark each fixed criterion completed again.",
"role": "user"
},
{
"content": "r3",
"role": "assistant"
},
{
"content": "Verification failed on these criteria:\n[ ] #1 le build passe\n\nFix the code so they pass. Address the notes precisely; do not touch what already passed. Mark each fixed criterion completed again.",
"role": "user"
}
],
"model": "loki",
"parallel_tool_calls": false,
"stream": true,
"stream_options": {
"include_usage": true
},
"temperature": 0.7,
"tools": [
{
"function": {
"description": "Run a shell command (bash syntax) and return stdout, stderr and exit code: inspect the system, read files and logs, run scripts. To CREATE or REWRITE a file use write instead — never echo/cat/python -c. Avoid destructive commands unless asked.",
"name": "bash",
"parameters": {
"properties": {
"command": {
"description": "The command",
"type": "string"
},
"timeout": {
"description": "Timeout s (default 30, max 300)",
"type": "integer"
}
},
"required": [
"command"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Create or replace a file with the exact content given (parent dirs created, content verbatim, no escaping). ALWAYS use this for a script or any text file — NEVER build one through the shell with echo, cat, python -c or Set-Content: quoting breaks.",
"name": "write",
"parameters": {
"properties": {
"content": {
"description": "Full content",
"type": "string"
},
"file": {
"description": "Path",
"type": "string"
}
},
"required": [
"file",
"content"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Patch a file by exact replacement: old → new. old must appear EXACTLY once (add context to make it unique). Prefer this over rewriting a whole file.",
"name": "edit",
"parameters": {
"properties": {
"file": {
"description": "Path",
"type": "string"
},
"new": {
"description": "Replacement",
"type": "string"
},
"old": {
"description": "Exact text to replace (unique)",
"type": "string"
}
},
"required": [
"file",
"old",
"new"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Bring back verbatim an earlier block archived when this conversation was compacted, from the id shown as recall:rN in the summary.",
"name": "recall",
"parameters": {
"properties": {
"id": {
"description": "Block id, e.g. \"r7\".",
"type": "string"
}
},
"required": [
"id"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Find an archived block and its recall id by keywords, when it is not listed in the current summary. Then call recall(id).",
"name": "recall_search",
"parameters": {
"properties": {
"limit": {
"description": "Default 5, max 20",
"type": "integer"
},
"query": {
"description": "Keywords describing the block.",
"type": "string"
}
},
"required": [
"query"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "List your scheduled tasks with their state, next run, and last result. Pass an id to get that task's full last report.",
"name": "task_list",
"parameters": {
"properties": {
"id": {
"description": "Show only this task, with its full last report",
"type": "string"
}
},
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Schedule work to run later, once or repeating. Give 'prompt' for an AI task, OR 'script' (a saved script name) for a script task that runs with no model and no tokens.",
"name": "task_create",
"parameters": {
"properties": {
"at": {
"description": "One-time task at this local time: \"HH:MM\" (next occurrence) or \"YYYY-MM-DD HH:MM\"",
"type": "string"
},
"enabled": {
"description": "Start active (default true)",
"type": "boolean"
},
"in_minutes": {
"description": "One-time task, N minutes from now",
"type": "integer"
},
"name": {
"type": "string"
},
"prompt": {
"description": "AI task: self-contained instruction to run each time (omit if using script)",
"type": "string"
},
"schedule": {
"description": "Repeating: \"@every 2h\" (interval), \"@every 1d@09:00\" (daily time), or a 5-field cron \"min hour dom mon dow\". Omit for a one-time task (use in_minutes or at)",
"type": "string"
},
"script": {
"description": "Script task: name of a script file you put in the scripts folder, run each time with no AI involved",
"type": "string"
}
},
"required": [
"name"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Change one or more fields of a scheduled task by id.",
"name": "task_update",
"parameters": {
"properties": {
"enabled": {
"description": "false pauses, true resumes",
"type": "boolean"
},
"id": {
"type": "string"
},
"name": {
"type": "string"
},
"prompt": {
"type": "string"
},
"schedule": {
"type": "string"
}
},
"required": [
"id"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Delete a scheduled task by id.",
"name": "task_delete",
"parameters": {
"properties": {
"id": {
"type": "string"
}
},
"required": [
"id"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Read a file with 1-indexed line numbers (offset/limit for long files). ALWAYS use this instead of cat/type: output is bounded and reading a file is REQUIRED before you may edit or overwrite it.",
"name": "read",
"parameters": {
"properties": {
"file": {
"description": "Path (relative to the working folder)",
"type": "string"
},
"limit": {
"description": "Lines (default 400, max 1000)",
"type": "integer"
},
"offset": {
"description": "Start line (default 1)",
"type": "integer"
}
},
"required": [
"file"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Search file contents with a Go regular expression. Returns path:line: text matches (bounded). path may be a file or a directory (default: whole working folder).",
"name": "grep",
"parameters": {
"properties": {
"glob": {
"description": "Only files matching this glob (e.g. *.go)",
"type": "string"
},
"path": {
"description": "File or directory to search (default .)",
"type": "string"
},
"pattern": {
"description": "Go regexp",
"type": "string"
}
},
"required": [
"pattern"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "List files matching a glob pattern (** supported, e.g. **/*.go), most recently modified first.",
"name": "glob",
"parameters": {
"properties": {
"pattern": {
"description": "Glob, e.g. src/**/*.ts",
"type": "string"
}
},
"required": [
"pattern"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Ask the user a clarifying question with 2-4 short options, then END your turn. Their answer arrives as the next user message. Use when a decision is genuinely theirs (scope, destructive change, ambiguous requirement) — not for things you can check yourself.",
"name": "ask",
"parameters": {
"properties": {
"options": {
"description": "2-4 short answer options",
"items": {
"type": "string"
},
"type": "array"
},
"question": {
"description": "The question",
"type": "string"
}
},
"required": [
"question"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Start a shell command in the BACKGROUND (dev server, watcher, long build) and return immediately with a job id. Output goes to a log file; check it with bash_tail. Use this instead of appending '&' in bash.",
"name": "bash_bg",
"parameters": {
"properties": {
"command": {
"description": "The command",
"type": "string"
}
},
"required": [
"command"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Read the last lines of a background job's output and its status (running/exited). kill:true stops the job.",
"name": "bash_tail",
"parameters": {
"properties": {
"id": {
"description": "Job id from bash_bg",
"type": "string"
},
"kill": {
"description": "Stop the job",
"type": "boolean"
},
"lines": {
"description": "Lines to show (default 40, max 200)",
"type": "integer"
}
},
"required": [
"id"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "git status of the working folder (branch + changed files, porcelain format).",
"name": "git_status",
"parameters": {
"properties": {
"dir": {
"description": "Repository subfolder (default: auto-detected)",
"type": "string"
}
},
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "git diff of the working folder (unstaged + staged). Optional file to diff a single path.",
"name": "git_diff",
"parameters": {
"properties": {
"dir": {
"description": "Repository subfolder (default: auto-detected)",
"type": "string"
},
"file": {
"description": "Limit the diff to this path",
"type": "string"
}
},
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Clone a git repository (https:// or git@) INTO the working folder of this discussion.",
"name": "git_clone",
"parameters": {
"properties": {
"dir": {
"description": "Target folder name (default: repo name)",
"type": "string"
},
"url": {
"description": "Repository URL",
"type": "string"
}
},
"required": [
"url"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Manage the acceptance criteria of the current task (the contract that defines DONE). action=add posits new criteria (texts[]), action=set updates one (id + status, optional note), action=list shows them, action=clear removes them all. Set criteria BEFORE building. Builder: mark each criterion completed once you have done AND checked it — verification only starts when nothing is left pending. Only the verification pass may mark passed or failed.",
"name": "criteria",
"parameters": {
"properties": {
"action": {
"enum": [
"add",
"set",
"list",
"clear"
],
"type": "string"
},
"id": {
"description": "set: criterion id",
"type": "integer"
},
"note": {
"description": "set: why it failed / how it was verified",
"type": "string"
},
"status": {
"enum": [
"pending",
"completed",
"passed",
"failed"
],
"type": "string"
},
"texts": {
"description": "add: one entry per criterion, short and testable",
"items": {
"type": "string"
},
"type": "array"
}
},
"required": [
"action"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Délègue une recherche ou une relecture à un rôle qui travaille dans son propre contexte et ne rend que sa réponse. Lecture seule.",
"name": "subagent",
"parameters": {
"properties": {
"role": {
"description": "explorer (trouver/cartographier), code-reviewer (relire un changement), planner (découper)",
"enum": [
"explorer",
"code-reviewer",
"planner"
],
"type": "string"
},
"task": {
"description": "La question, précise et autonome : le rôle ne voit pas la discussion.",
"type": "string"
}
},
"required": [
"role",
"task"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Photographie une page web (JS exécuté) pour la MONTRER. La réponse donne la ligne markdown à recopier. Tu ne vois pas l'image.",
"name": "web_screenshot",
"parameters": {
"properties": {
"full_page": {
"description": "Page entière (lourd). Défaut false.",
"type": "boolean"
},
"url": {
"description": "URL complète",
"type": "string"
}
},
"required": [
"url"
],
"type": "object"
}
},
"type": "function"
}
]
},
{
"messages": [
{
"content": "You are the VERIFICATION pass of code mode. You independently check the acceptance criteria against the actual code — you did not write it, trust nothing.\n\nFor each criterion still pending or failed:\n1. Understand what it actually requires.\n2. Check the real changes: git_diff, read the touched files, run the relevant build/test command with bash when applicable.\n3. Record the verdict with the criteria tool: action=set, status=passed (satisfied) or failed (with a note saying exactly what is missing — actionable, one line).\n\nRules:\n- Trivial or non-code criteria that are plainly satisfied: pass them without exploring.\n- Only fail a criterion that genuinely misses its requirement; a different-but-valid implementation passes.\n- Do NOT fix anything yourself. You only verify and report.\n- Be fast: check the diff first, read only what the criteria require.\n\nMachine: host={{HOST}}, cwd={{TMP}}/workspace/discussions/{{CONV}}. This is your working folder: relative paths in write/edit/bash resolve here, and it is the DEFAULT place for scratch work — notes, outputs, a clone, a test. But it is DISPOSABLE: deleting the discussion wipes it. Any script you want to KEEP (or schedule), write it into your scripts folder {{TMP}}/scripts instead — a separate folder a workspace wipe won't touch; you write and run scripts there normally. Do NOT install or write files into system directories such as /usr/local/bin, /usr, /bin or /etc: those need root and are not yours. Only use an absolute path outside this folder (except your scripts folder) when the user explicitly named that location.",
"role": "system"
},
{
"content": "Task:\nVerification failed on these criteria:\n[ ] #1 le build passe\n\nFix the code so they pass. Address the notes precisely; do not touch what already passed. Mark each fixed criterion completed again.\n\nAcceptance criteria:\n[ ] #1 le build passe\n\nVerify each non-passed criterion now (git_diff first), and record every verdict with the criteria tool.",
"role": "user"
}
],
"model": "loki",
"parallel_tool_calls": false,
"stream": true,
"stream_options": {
"include_usage": true
},
"temperature": 0.7,
"tools": [
{
"function": {
"description": "Run a shell command (bash syntax) and return stdout, stderr and exit code: inspect the system, read files and logs, run scripts. To CREATE or REWRITE a file use write instead — never echo/cat/python -c. Avoid destructive commands unless asked.",
"name": "bash",
"parameters": {
"properties": {
"command": {
"description": "The command",
"type": "string"
},
"timeout": {
"description": "Timeout s (default 30, max 300)",
"type": "integer"
}
},
"required": [
"command"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Read a file with 1-indexed line numbers (offset/limit for long files). ALWAYS use this instead of cat/type: output is bounded and reading a file is REQUIRED before you may edit or overwrite it.",
"name": "read",
"parameters": {
"properties": {
"file": {
"description": "Path (relative to the working folder)",
"type": "string"
},
"limit": {
"description": "Lines (default 400, max 1000)",
"type": "integer"
},
"offset": {
"description": "Start line (default 1)",
"type": "integer"
}
},
"required": [
"file"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Search file contents with a Go regular expression. Returns path:line: text matches (bounded). path may be a file or a directory (default: whole working folder).",
"name": "grep",
"parameters": {
"properties": {
"glob": {
"description": "Only files matching this glob (e.g. *.go)",
"type": "string"
},
"path": {
"description": "File or directory to search (default .)",
"type": "string"
},
"pattern": {
"description": "Go regexp",
"type": "string"
}
},
"required": [
"pattern"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "List files matching a glob pattern (** supported, e.g. **/*.go), most recently modified first.",
"name": "glob",
"parameters": {
"properties": {
"pattern": {
"description": "Glob, e.g. src/**/*.ts",
"type": "string"
}
},
"required": [
"pattern"
],
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "git status of the working folder (branch + changed files, porcelain format).",
"name": "git_status",
"parameters": {
"properties": {
"dir": {
"description": "Repository subfolder (default: auto-detected)",
"type": "string"
}
},
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "git diff of the working folder (unstaged + staged). Optional file to diff a single path.",
"name": "git_diff",
"parameters": {
"properties": {
"dir": {
"description": "Repository subfolder (default: auto-detected)",
"type": "string"
},
"file": {
"description": "Limit the diff to this path",
"type": "string"
}
},
"type": "object"
}
},
"type": "function"
},
{
"function": {
"description": "Manage the acceptance criteria of the current task (the contract that defines DONE). action=add posits new criteria (texts[]), action=set updates one (id + status, optional note), action=list shows them, action=clear removes them all. Set criteria BEFORE building. Builder: mark each criterion completed once you have done AND checked it — verification only starts when nothing is left pending. Only the verification pass may mark passed or failed.",
"name": "criteria",
"parameters": {
"properties": {
"action": {
"enum": [
"add",
"set",
"list",
"clear"
],
"type": "string"
},
"id": {
"description": "set: criterion id",
"type": "integer"
},
"note": {
"description": "set: why it failed / how it was verified",
"type": "string"
},
"status": {
"enum": [
"pending",
"completed",
"passed",
"failed"
],
"type": "string"
},
"texts": {
"description": "add: one entry per criterion, short and testable",
"items": {
"type": "string"
},
"type": "array"
}
},
"required": [
"action"
],
"type": "object"
}
},
"type": "function"
}
]
}
],
"log": [
{
"user": "Construis le jeu"
},
{
"content": "r0",
"seq0": 2,
"toks": 1
},
{
"stats": "…"
},
{
"stats": "…"
},
{
"stats": "…"
},
{
"role": "builder"
},
{
"content": "r1",
"seq0": 7,
"toks": 1
},
{
"stats": "…"
},
{
"stats": "…"
},
{
"stats": "…"
},
{
"role": ""
},
{
"role": "builder"
},
{
"content": "r2",
"seq0": 13,
"toks": 1
},
{
"stats": "…"
},
{
"stats": "…"
},
{
"stats": "…"
},
{
"role": ""
},
{
"role": "verifier"
},
{
"content": "v1",
"seq0": 19,
"toks": 1
},
{
"ctx_isolated": true,
"stats": "…"
},
{
"ctx_isolated": true,
"stats": "…"
},
{
"ctx_isolated": true,
"stats": "…"
},
{
"role": ""
},
{
"role": "builder"
},
{
"content": "r3",
"seq0": 25,
"toks": 1
},
{
"stats": "…"
},
{
"stats": "…"
},
{
"stats": "…"
},
{
"role": ""
},
{
"role": "verifier"
},
{
"content": "v2",
"seq0": 31,
"toks": 1
},
{
"ctx_isolated": true,
"stats": "…"
},
{
"ctx_isolated": true,
"stats": "…"
},
{
"ctx_isolated": true,
"stats": "…"
},
{
"role": ""
},
{
"role": "builder"
},
{
"content": "r4",
"seq0": 37,
"toks": 1
},
{
"stats": "…"
},
{
"stats": "…"
},
{
"stats": "…"
},
{
"role": ""
},
{
"role": "verifier"
},
{
"content": "v3",
"seq0": 43,
"toks": 1
},
{
"ctx_isolated": true,
"stats": "…"
},
{
"ctx_isolated": true,
"stats": "…"
},
{
"ctx_isolated": true,
"stats": "…"
},
{
"role": ""
},
{
"verify_done": "gave_up"
},
{
"content": "\n\n_(vérification : 1 critère(s) non satisfaits après 2 corrections — voir le panneau des critères)_",
"seq0": 49,
"toks": 1
},
{
"turn_done": true
}
],
"messages": [
{
"role": "user",
"content": "Construis le jeu"
},
{
"role": "assistant",
"content": "r0"
},
{
"role": "user",
"content": "Not done yet — these acceptance criteria are still open:\n[ ] #1 le build passe\n\nContinue working on them. Once one is done and checked, mark it completed (criteria action=set, status=completed). If you need the user's decision, use ask."
},
{
"role": "assistant",
"content": "r1"
},
{
"role": "user",
"content": "Not done yet — these acceptance criteria are still open:\n[ ] #1 le build passe\n\nContinue working on them. Once one is done and checked, mark it completed (criteria action=set, status=completed). If you need the user's decision, use ask."
},
{
"role": "assistant",
"content": "r2"
},
{
"role": "user",
"content": "Verification failed on these criteria:\n[ ] #1 le build passe\n\nFix the code so they pass. Address the notes precisely; do not touch what already passed. Mark each fixed criterion completed again."
},
{
"role": "assistant",
"content": "r3"
},
{
"role": "user",
"content": "Verification failed on these criteria:\n[ ] #1 le build passe\n\nFix the code so they pass. Address the notes precisely; do not touch what already passed. Mark each fixed criterion completed again."
},
{
"role": "assistant",
"content": "r4"
}
],
"status": 200
}
}
Reference in new issue View Git Blame Copy Permalink