…esis

Based on analysis of 35+ sources including JetBrains, Anthropic, SWE-bench
leaderboard, and Wink's 42k-trajectory production study.

## Changes

**Doom-loop fingerprinting (loopDetectionService + agent-core)**
Replace consecutive-only threshold=5 with a sliding 20-call window where
any fingerprint appearing 3+ times = doom loop. Catches non-consecutive
repetition patterns that the old approach missed (the #1 recovery category
in production, 39% of all interventions per Wink).

**Silent sensors (baselineCheck + postEditVerify)**
Remove PASSED output from baseline verify — silent on pass preserves context
budget. Structured remediation steps on failure: read error → fix root cause
→ re-run command.

**Read-only plan subagent (builtin-agents)**
Add `plan` builtin agent with write tools structurally absent from its
schema. Prevents the "accidental edit during planning" failure mode that
every successful harness independently converged on.

**Checkpoint commits (agentCore + agent-core)**
Add `gitSnapshotBeforeEdit()` — creates a named shadow-repo commit before
every file-mutating tool call in the AgentCore path. Durable across crashes,
fire-and-forget so it never blocks tool execution. Pairs with existing
in-memory CheckpointStore for dual-layer rollback.

**Scope lock (scopeLock service + agent-core)**
New `ScopeLockService` singleton — activated from a sprint contract with a
permitted file set. Any write outside the set is intercepted before the tool
executes, returning a structured violation message and blocking the edit.
Addresses the 6.62% "unrequested changes" failure category.

**Observation masking (chatCompressionService + agent-core)**
Add `applyObservationMask()` — replaces old tool call/result pairs with a
placeholder, keeping the last N verbatim. Applied before LLM compression in
the AgentCore compaction path. JetBrains (2025): observation masking reduces
peak tokens 26-54% while LLM summarisation made agents run 15% longer.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

… example shape to description (#177)

Reported failure mode: smaller / older models (Qwen3, Minimax variants,
some open-weights routes) JSON-encode the entire `questions` array as a
string instead of emitting it as a native array literal. validateToolParams
threw with "Parameter \"questions\" must be an array" — useless feedback
since the model HAD sent the array, just stringified.

Three changes, layered:

1. Silent coercion in validateToolParams. If `questions` is a string that
   parses as JSON, parse it and continue. Logs a debugLogger.warn so the
   signal stays visible — silent coercion would mask a real upstream
   regression if model behavior shifts. Catches ~all of the user-reported
   failures without a retry round-trip.

2. Example shape added to the tool description. Models replicate concrete
   examples better than they synthesize from abstract schemas with 3
   levels of nesting. Placeholder text is clearly labeled as shape-only
   so models don't cargo-cult the example values into their actual
   questions.

3. Sharper error message for the residual case (non-JSON garbage in the
   string slot): "Pass `questions` as a real array literal, not a
   JSON-encoded string." Clear, specific, tells the model exactly which
   strategy to drop.

Considered but rejected:
- Schema relaxation (allow `options: string[]`, default multiSelect):
  API change, breaks downstream `option.description` consumers (dialog
  UI, ACP renderer), premature without data showing #1+#2 are
  insufficient.

Tests updated with two new cases: stringified-array coercion, non-JSON
string error path. 5319 core / 0 fail; lint + typecheck clean.

Co-authored-by: Automaker <automaker@localhost>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>