Journal

Journal — cross-round memory

A markdown file (the caller passes its path as memo_path). Each evolve appends one ## Round N — <label> section. The file is the meta-agent's only memory across rounds — there is no persistent conversation, and the orchestrator fills in gating outcomes between rounds so the next agent can see which past bets paid off.

The orchestrator pre-renders a machine index at _meta_scratch/CONTEXT.md (when the journal has prior entries): lever scoreboard, recent hypotheses, reverted bets. Read that first — it's a cheap way to see what's been tried without scanning the full journal prose.

On entry — read once

1. Read <memo_path> — full cross-round history, prose included.
2. Read _meta_scratch/CONTEXT.md (when present) — compact index of lever usage, attribution, and reverted hypotheses.

Look for:

• Hypotheses with gating_outcome: reverted → do not re-propose without new evidence (hard invariant #4 variant).
• The lever scoreboard's prediction hit rate → if Action landed 1/12 times while Instruction landed 4/6, update your prior.
• Outstanding needs_from_human items (free-prose; scan for them).

On exit — append one section

Required shape:

## Round N — <short label in 2-5 words>

<!-- journal:frontmatter
round: N
timestamp: 2026-04-27T15:30:00Z    # ISO-8601 UTC; orchestrator can override
hypothesis_id: h_<short slug>      # stable across related rounds
levers: [action]                   # subset of {configuration, control, action, instruction}
predicted_affected: [task_a, task_b]
gating_outcome: pending             # orchestrator back-fills next round
gating_attribution: pending         # orchestrator back-fills next round
expected_global_gain: "<one-line cluster-level upside>"
regression_risk: "<main risk outside predicted_affected>"
cost_shift: "<expected token/cost delta or direction>"
rollback_trigger: "<observable signal to revert next round>"
-->

### Why

One paragraph on the pattern that motivated the round — the
cluster of trajectories and the shape they share.

### Changes

- `tools/<name>.py` — new @tool, one-line signature description
- `config.yaml` — register above under `tools.custom`

### Evidence

- `task_a` frontmatter: `judge_missing_capability.summary: "<gap phrase>"`
- `task_a` step <n> body: `<40-char quote showing the failure shape>`
- `task_b` exit_reason=max_steps, <n>/<max_steps> spent on the same
  pattern

### Uncertainty

One-liner on what might go wrong and how you'd know.

Frontmatter keys — all required

• round (int) — monotonic round index.
• timestamp (ISO-8601 UTC string) — when appended.
• hypothesis_id (string slug) — short, stable across related rounds. Prefer semantic ids that name the intervention shape (e.g. h_<cluster>_v1, h_<tool_name>_retry_v2) over auto-generated ones. If a later round extends the same hypothesis, reuse the same id with a version suffix.
• levers (list) — subset of {configuration, control, action, instruction}.
• predicted_affected (list of task_ids) — the task ids you claim this round's change should flip from F to T (for corrective / preservative-transfer candidates) or protect from T→F regression (for preservative-lock candidates). Lists with zero entries are legitimate when the claim is about cost, not pass-rate. Preservative-lock candidates can list their currently-passing cluster here to make the "don't regress these" claim explicit. Over-predicting hurts your attribution score; the orchestrator grades this.
• gating_outcome (string) — pending on first write. The orchestrator fills with accepted / reverted / noop after the benchmark round runs.
• gating_attribution (dict | string) — pending on first write. The orchestrator fills with {task_id: flipped|still_F|regressed|still_T} after evaluation.
• expected_global_gain (string) — one-line claim of global upside (cluster-level, not single-task anecdote). For corrective / preservative-transfer candidates: "flips N tasks in cluster X". For preservative-lock candidates: "prevents expected regression of N tasks if future rounds drift" — this is a legitimate gain even with zero flips.
• regression_risk (string) — primary collateral risk outside predicted_affected.
• cost_shift (string) — expected cost/tokens movement if this change lands.
• rollback_trigger (string) — concrete signal that means "revert this hypothesis" next round.

Frontmatter keys — required conditionally

• cited_candidates (list of strings) — required when the round's config changes (the orchestrator's evidence gate enforces it). Lists the C-NNN IDs from _meta_scratch/candidates.md that motivated this round's changes. Example: cited_candidates: [C-001, C-004]. See the analyze skill for the candidates.md format. Omit entirely on noop rounds (byte- identical config copy) — they don't need evidence.
• retry_rationale (string) — required only if your (levers, predicted_affected) signature matches a reverted round. One-line explanation of what evidence is new since the last try, e.g. retry_rationale: "R3 tried instruction-only; this round found tool_error_counts spike for WebFetch that the prior attempt ignored". Skip when your signature is novel. See the novelty findings file (_meta_scratch/NOVELTY_FAIL.md) if the gate rejects your round — it names the colliding prior round.

Prose body — four suggested headings, not enforced

The parser ignores the prose. These headings exist for consistency:

• ### Why — the cluster of trajectories and the pattern they share. For corrective rounds: the failure shape. For preservative-lock rounds: the habit present in the passing cluster that would be at risk without the encoding. For preservative-transfer rounds: the habit fired on the passing cluster + absent on the failing cluster.
• ### Changes — bulleted list, one bullet per authored file or config edit.
• ### Evidence — quoted frontmatter or body excerpts with task_id citations. Hand-wavy "several tasks did X" is not evidence. Preservative-transfer rounds must cite both ends (passing cluster body + failing cluster body), consistent with candidates.md.
• ### Uncertainty — what might fail and how you'd tell.

Include the same four signals in prose when stakes are high: expected_global_gain, regression_risk, cost_shift, rollback_trigger. This makes Pareto-style tradeoffs explicit across rounds instead of implicit in vague narratives.

Write more when the round is high-stakes, less when it's a small config nudge. Don't abbreviate the frontmatter — that's what the orchestrator reads.

Rules

• Append, never edit prior sections. The orchestrator DOES edit prior sections — it back-fills gating_outcome and gating_attribution surgically without touching prose. You, as the agent, never rewrite an old round.
• Cite task ids + field or body quote in Evidence. Vague is bad.
• Do not re-propose reverted hypotheses without a specific reason the prior attempt was buggy (not "the idea was right but my implementation wasn't" — come back with a different shape, a different lever, or more evidence).
• Multi-lever rounds are allowed when the evidence warrants (see analyze skill → Multi-lever rounds). List each change under its own bullet in Changes and cite one candidate ID per change in cited_candidates. Attribution is per-round, so bundling does widen the ambiguity when gating — that's a reason to keep each change independently supported (own candidate, own cluster, own retroactive-check variant), not a reason to drop to a single change per round.

Legacy compat

Entries without the <!-- journal:frontmatter ... --> block are skipped by the machine parser but still render as prose for readers. Early rounds that predate the journal format will be invisible in CONTEXT.md's scoreboard — that's fine, new rounds start fresh.

Example minimal append

Small config tune, one gap, one lever. Because this changes the config (CostGuardProcessor kwargs), the entry must cite a candidate from _meta_scratch/candidates.md:

## Round 3 — loosen cost guard

<!-- journal:frontmatter
round: 3
timestamp: 2026-04-27T10:12:00Z
hypothesis_id: h_cost_guard_v1
levers: [configuration]
predicted_affected: [task_42]
cited_candidates: [C-001]
gating_outcome: pending
gating_attribution: pending
expected_global_gain: "Should reduce budget_exceeded cluster on long-horizon tasks"
regression_risk: "May increase cost on easy tasks that already pass"
cost_shift: "+5-10% tokens expected on median run"
rollback_trigger: "If pass_rate is flat/down and cost rises >10%, revert"
-->

### Why

task_42 hit budget_exceeded at step 18/30 while still productively
calling tools. CostGuardProcessor max_usd=0.5 is too tight for
Level-3 tasks that require >20 WebFetch calls.

### Changes

- `config.yaml` — CostGuardProcessor max_usd 0.5 → 1.0

### Evidence

- `task_42` frontmatter: `exit_reason: budget_exceeded`, cost_usd: 0.51
- `task_42` step 18 body: still making progress ("Found the relevant
  Wikipedia section — let me fetch one more to confirm").

### Uncertainty

Looser cap may push the per-round bill up ~10%. If R4 pass_rate
drops while cost climbs, revert.