Agent context budget

agent-context-budget

Context governor for multi-agent rounds. The core rule is simple: keep .coord/ as the canonical state, pass agents compact packets, and never paste raw logs or unbounded memory into the main session.

When to Use

Use this before or during:

• Large Codex + Gemini + Claude runs.
• Cross-session resumes where the prior conversation is too large.
• Any round with more than two delegate tasks.
• Any workflow where .coord/memory.yml, .ai/*_log_*.txt, or agent summaries are starting to dominate the prompt.

Not for small one-agent edits. Use the direct delegate skill instead.

Default Policy

If .coord/plan.yml lacks context_policy, add this block:

context_policy:
  main_session_token_budget: 3000
  task_packet_token_budget: 6000
  result_summary_word_budget: 250
  memory_digest_token_budget: 1200
  log_tail_lines_on_error: 50
  raw_log_policy: path-only
  agentmemory: optional

  # W3 — $ cost gate per task (v0.2.2+)
  # Optional. If set, agent-acceptance-gate flags any task whose
  # codex/gemini delegate exceeded the cap. Estimated from token
  # usage × provider price (Anthropic / OpenAI / Google rates).
  # Set per-task in plan.yml tasks[].budget.max_cost_usd to override.
  default_max_cost_usd: 0.50  # default $ ceiling per task
  total_round_max_cost_usd: 5.00  # hard stop for the entire round

Why both token and cost budgets? Token gate prevents context bloat (a session quality concern). Cost gate prevents runaway delegation spend (a financial concern). They're orthogonal: a 2k-token delegate call can cost $0.05 (Claude Haiku) or $0.50 (Claude Opus), so token count alone doesn't bound dollars.

Per-task override example:

# in plan.yml
tasks:
  - id: T1
    agent: codex
    slug: simple-refactor
    budget:
      max_cost_usd: 0.10  # this task is mechanical, cap low
  - id: T2
    agent: codex
    slug: complex-rewrite
    budget:
      max_cost_usd: 2.00  # this task needs frontier model, allow higher

Workflow

1. Read .coord/plan.yml and .coord/memory.yml if present.
2. Write .coord/context_<NNN>.md with:
   • round goal and success criteria
   • task graph and write ownership
   • per-agent context packets to include
   • context intentionally excluded
   • optional agentmemory recall queries, if available
3. Write or refresh .coord/session_primer.md with:
   • current decisions
   • open questions
   • active round status
   • recent artifacts by path
   • no raw logs
4. Enforce result packets:
   • delegate summaries are <= 250 words
   • changed files are listed by path
   • tests are listed by command + result
   • risks are explicit and short
5. If a log is needed, include only the path. Read the last 50 lines only when the corresponding result.json status is error.

Optional agentmemory

agentmemory is a recall cache, not the source of truth.

• Query it only to enrich .coord/session_primer.md.
• Store only compact memory candidates: accepted decisions, resolved open questions, artifact summaries, and final session outcomes.
• If it is unavailable, continue with .coord/memory.yml only.
• Never use vector recall as acceptance evidence. The acceptance gate must read .coord/plan.yml, result files, reconciliation, and tests.

Memory Promotion Rules

Promote only:

• Decisions with a one-sentence what and why.
• Open questions with a blocker and suggested next owner.
• Artifact pointers with one-line summaries.
• Agent session outcomes with status and result-summary path.

Do not promote:

• Raw logs.
• Full diffs.
• Source code.
• Long analysis.
• Secrets or credentials.

Output

End with:

[agent-context-budget]
  Round: <N>
  Policy: .coord/plan.yml context_policy
  Context plan: .coord/context_<NNN>.md
  Session primer: .coord/session_primer.md
  Raw logs: path-only; failure tail max 50 lines
  agentmemory: optional cache, not required

Common Mistakes

• Pasting full logs into chat. Store the path; read bounded tail only on failure.
• Treating memory as a transcript. Promote decisions, questions, artifacts, and outcomes only.
• Giving Gemini only .ai/ paths. Inline critical context because .ai/ may be gitignored.
• Using agentmemory as the gate. Gate with tests and .coord/ artifacts.

Subagent review (keep main session lean)

When: ≥ 3 task packets exist and you want to verify they stayed within the declared task_packet_token_budget (default 6000 tokens ≈ 24 KB) without re-reading them all in the main session.

Why: After writing .coord/context_<NNN>.md + session_primer.md, the main session has the policy in context but doesn't need to re-read every packet to verify compliance. A subagent can scan all packets and return only the over-budget list.

Pattern:

Spawn `code-reviewer` subagent:
- Read .coord/context_<NNN>.md (the declared per-task budgets)
- Read each .ai/<agent>_task_<NNN>_*.md file referenced in plan.yml
- For each, count rough tokens (word count × 1.3 as approximation)
- Return: a single PASS/FAIL line + list of any task slugs > 120%
  of declared budget + suggested compression target

Main session reads only the verdict.

If subagent reports FAIL, regenerate the over-budget packets with tighter prompts before invoking delegates.

Commit Boundary

Every agent boundary is a commit boundary (see global rule: ~/.claude/CLAUDE.md → "Commit Discipline for Multi-Agent Work").

Specific to this skill: writing .coord/context_<NNN>.md and .coord/session_primer.md are session-setup artifacts — commit them as a single round-prep commit so the budget policy used for the round is auditable. Pattern:

git add .coord/context_<NNN>.md .coord/session_primer.md
git commit -m "context: round <N> budget plan + session primer"

The acceptance gate later verifies actual round consumption against this committed policy.