feat(headless): feed tool failure traces to RSI prompt loop

[Astro-Han]

What

Adds the first Issue #64 follow-up: feed aggregated tool_failed run-trace signal into the RSI meta-agent input.

The v1 RSI loop in #149 validated the structure, but the real unattended run discarded most candidates for coverage_regressed. This PR gives the prompt-editing meta-agent a compact round-level signal about which tools failed repeatedly, so it is no longer editing only from scalar pass/fail outcomes and the last two tool calls.

Pieces

• Trace artifact plumbing: the Harbor task runner derives the host trace sidecar path from runtimeRefs and preserves it in fixed-prompt WAL outputs, while HarborCellOutput stays limited to the cell artifact contract.
• Trajectory digest enrichment: extractTrajectoryDigest reads the trace ledger when available, joins tool_failed events with model-visible tool-call arg keys, and emits bounded toolFailures entries.
• Prompt hardening: trace-derived fields are reduced to prompt-safe tokens before rendering, malformed trace sidecars fail loudly, and the tool-failure signal appears only in the aggregate summary instead of being duplicated in digest JSON.
• Compatibility boundary: existing runtimeEventsPath behavior remains unchanged; reward-hack scanning still uses model-visible runtime events, not the diagnostic trace ledger.

Validation

• npm run -w @maka/headless test — 405 tests pass
• npm run -w @maka/headless typecheck
• git diff --check

Follow-ups (not in this PR)

• A/B harness to compare a hand-authored / opencode-derived baseline prompt against the current seed using the existing acceptance policy.
• Real unattended RSI run with this signal enabled, after review, to see whether coverage-regression discards decrease.

Part of #64.

[Astro-Han]

What

Adds the first Issue #64 follow-up: feed aggregated tool_failed run-trace signal into the RSI meta-agent input.

The v1 RSI loop in #149 validated the structure, but the real unattended run discarded most candidates for coverage_regressed. This PR gives the prompt-editing meta-agent a compact round-level signal about which tools failed repeatedly, so it is no longer editing only from scalar pass/fail outcomes and the last two tool calls.

Pieces

• Trace artifact plumbing: the Harbor task runner derives the host trace sidecar path from runtimeRefs and preserves it in fixed-prompt WAL outputs, while HarborCellOutput stays limited to the cell artifact contract.
• Trajectory digest enrichment: extractTrajectoryDigest reads the trace ledger when available, joins tool_failed events with model-visible tool-call arg keys, and emits bounded toolFailures entries.
• Prompt hardening: trace-derived fields are reduced to prompt-safe tokens before rendering, malformed trace sidecars fail loudly, and the tool-failure signal appears only in the aggregate summary instead of being duplicated in digest JSON.
• Compatibility boundary: existing runtimeEventsPath behavior remains unchanged; reward-hack scanning still uses model-visible runtime events, not the diagnostic trace ledger.

Validation

• npm run -w @maka/headless test — 405 tests pass
• npm run -w @maka/headless typecheck
• git diff --check

Follow-ups (not in this PR)

• A/B harness to compare a hand-authored / opencode-derived baseline prompt against the current seed using the existing acceptance policy.
• Real unattended RSI run with this signal enabled, after review, to see whether coverage-regression discards decrease.

Part of #64.

[Astro-Han]

What

Adds the first Issue #64 follow-up: feed aggregated tool_failed run-trace signal into the RSI meta-agent input.

The v1 RSI loop in #149 validated the structure, but the real unattended run discarded most candidates for coverage_regressed. This PR gives the prompt-editing meta-agent a compact round-level signal about which tools failed repeatedly, so it is no longer editing only from scalar pass/fail outcomes and the last two tool calls.

Pieces

• Trace artifact plumbing: the Harbor task runner derives the host trace sidecar path from runtimeRefs and preserves it in fixed-prompt WAL outputs, while HarborCellOutput stays limited to the cell artifact contract.
• Trajectory digest enrichment: extractTrajectoryDigest reads the trace ledger when available, joins tool_failed events with model-visible tool-call arg keys, and emits bounded toolFailures entries.
• Prompt hardening: trace-derived fields are reduced to prompt-safe tokens before rendering, malformed trace sidecars fail loudly, and the tool-failure signal appears only in the aggregate summary instead of being duplicated in digest JSON.
• Compatibility boundary: existing runtimeEventsPath behavior remains unchanged; reward-hack scanning still uses model-visible runtime events, not the diagnostic trace ledger.

Validation

• npm run -w @maka/headless test — 405 tests pass
• npm run -w @maka/headless typecheck
• git diff --check

Follow-ups (not in this PR)

• A/B harness to compare a hand-authored / opencode-derived baseline prompt against the current seed using the existing acceptance policy.
• Real unattended RSI run with this signal enabled, after review, to see whether coverage-regression discards decrease.

Part of #64.

[Astro-Han]

What

Adds the first Issue #64 follow-up: feed aggregated tool_failed run-trace signal into the RSI meta-agent input.

The v1 RSI loop in #149 validated the structure, but the real unattended run discarded most candidates for coverage_regressed. This PR gives the prompt-editing meta-agent a compact round-level signal about which tools failed repeatedly, so it is no longer editing only from scalar pass/fail outcomes and the last two tool calls.

Pieces

• Trace artifact plumbing: the Harbor task runner derives the host trace sidecar path from runtimeRefs and preserves it in fixed-prompt WAL outputs, while HarborCellOutput stays limited to the cell artifact contract.
• Trajectory digest enrichment: extractTrajectoryDigest reads the trace ledger when available, joins tool_failed events with model-visible tool-call arg keys, and emits bounded toolFailures entries.
• Prompt hardening: trace-derived fields are reduced to prompt-safe tokens before rendering, malformed trace sidecars fail loudly, and the tool-failure signal appears only in the aggregate summary instead of being duplicated in digest JSON.
• Compatibility boundary: existing runtimeEventsPath behavior remains unchanged; reward-hack scanning still uses model-visible runtime events, not the diagnostic trace ledger.

Validation

• npm run -w @maka/headless test — 405 tests pass
• npm run -w @maka/headless typecheck
• git diff --check

Follow-ups (not in this PR)

• A/B harness to compare a hand-authored / opencode-derived baseline prompt against the current seed using the existing acceptance policy.
• Real unattended RSI run with this signal enabled, after review, to see whether coverage-regression discards decrease.

Part of #64.

[Astro-Han]

What

Adds the first Issue #64 follow-up: feed aggregated tool_failed run-trace signal into the RSI meta-agent input.

The v1 RSI loop in #149 validated the structure, but the real unattended run discarded most candidates for coverage_regressed. This PR gives the prompt-editing meta-agent a compact round-level signal about which tools failed repeatedly, so it is no longer editing only from scalar pass/fail outcomes and the last two tool calls.

Pieces

• Trace artifact plumbing: the Harbor task runner derives the host trace sidecar path from runtimeRefs and preserves it in fixed-prompt WAL outputs, while HarborCellOutput stays limited to the cell artifact contract.
• Trajectory digest enrichment: extractTrajectoryDigest reads the trace ledger when available, joins tool_failed events with model-visible tool-call arg keys, and emits bounded toolFailures entries.
• Prompt hardening: trace-derived fields are reduced to prompt-safe tokens before rendering, malformed trace sidecars fail loudly, and the tool-failure signal appears only in the aggregate summary instead of being duplicated in digest JSON.
• Compatibility boundary: existing runtimeEventsPath behavior remains unchanged; reward-hack scanning still uses model-visible runtime events, not the diagnostic trace ledger.

Validation

• npm run -w @maka/headless test — 405 tests pass
• npm run -w @maka/headless typecheck
• git diff --check

Follow-ups (not in this PR)

• A/B harness to compare a hand-authored / opencode-derived baseline prompt against the current seed using the existing acceptance policy.
• Real unattended RSI run with this signal enabled, after review, to see whether coverage-regression discards decrease.

Part of #64.

[Astro-Han]

What

Adds the first Issue #64 follow-up: feed aggregated tool_failed run-trace signal into the RSI meta-agent input.

The v1 RSI loop in #149 validated the structure, but the real unattended run discarded most candidates for coverage_regressed. This PR gives the prompt-editing meta-agent a compact round-level signal about which tools failed repeatedly, so it is no longer editing only from scalar pass/fail outcomes and the last two tool calls.

Pieces

• Trace artifact plumbing: the Harbor task runner derives the host trace sidecar path from runtimeRefs and preserves it in fixed-prompt WAL outputs, while HarborCellOutput stays limited to the cell artifact contract.
• Trajectory digest enrichment: extractTrajectoryDigest reads the trace ledger when available, joins tool_failed events with model-visible tool-call arg keys, and emits bounded toolFailures entries.
• Prompt hardening: trace-derived fields are reduced to prompt-safe tokens before rendering, malformed trace sidecars fail loudly, and the tool-failure signal appears only in the aggregate summary instead of being duplicated in digest JSON.
• Compatibility boundary: existing runtimeEventsPath behavior remains unchanged; reward-hack scanning still uses model-visible runtime events, not the diagnostic trace ledger.

Validation

• npm run -w @maka/headless test — 405 tests pass
• npm run -w @maka/headless typecheck
• git diff --check

Follow-ups (not in this PR)

• A/B harness to compare a hand-authored / opencode-derived baseline prompt against the current seed using the existing acceptance policy.
• Real unattended RSI run with this signal enabled, after review, to see whether coverage-regression discards decrease.

Part of #64.

[Astro-Han]

What

Adds the first Issue #64 follow-up: feed aggregated tool_failed run-trace signal into the RSI meta-agent input.

The v1 RSI loop in #149 validated the structure, but the real unattended run discarded most candidates for coverage_regressed. This PR gives the prompt-editing meta-agent a compact round-level signal about which tools failed repeatedly, so it is no longer editing only from scalar pass/fail outcomes and the last two tool calls.

Pieces

• Trace artifact plumbing: the Harbor task runner derives the host trace sidecar path from runtimeRefs and preserves it in fixed-prompt WAL outputs, while HarborCellOutput stays limited to the cell artifact contract.
• Trajectory digest enrichment: extractTrajectoryDigest reads the trace ledger when available, joins tool_failed events with model-visible tool-call arg keys, and emits bounded toolFailures entries.
• Prompt hardening: trace-derived fields are reduced to prompt-safe tokens before rendering, malformed trace sidecars fail loudly, and the tool-failure signal appears only in the aggregate summary instead of being duplicated in digest JSON.
• Compatibility boundary: existing runtimeEventsPath behavior remains unchanged; reward-hack scanning still uses model-visible runtime events, not the diagnostic trace ledger.

Validation

• npm run -w @maka/headless test — 405 tests pass
• npm run -w @maka/headless typecheck
• git diff --check

Follow-ups (not in this PR)

• A/B harness to compare a hand-authored / opencode-derived baseline prompt against the current seed using the existing acceptance policy.
• Real unattended RSI run with this signal enabled, after review, to see whether coverage-regression discards decrease.

Part of #64.

[Astro-Han]

What

Adds the first Issue #64 follow-up: feed aggregated tool_failed run-trace signal into the RSI meta-agent input.

The v1 RSI loop in #149 validated the structure, but the real unattended run discarded most candidates for coverage_regressed. This PR gives the prompt-editing meta-agent a compact round-level signal about which tools failed repeatedly, so it is no longer editing only from scalar pass/fail outcomes and the last two tool calls.

Pieces

• Trace artifact plumbing: the Harbor task runner derives the host trace sidecar path from runtimeRefs and preserves it in fixed-prompt WAL outputs, while HarborCellOutput stays limited to the cell artifact contract.
• Trajectory digest enrichment: extractTrajectoryDigest reads the trace ledger when available, joins tool_failed events with model-visible tool-call arg keys, and emits bounded toolFailures entries.
• Prompt hardening: trace-derived fields are reduced to prompt-safe tokens before rendering, malformed trace sidecars fail loudly, and the tool-failure signal appears only in the aggregate summary instead of being duplicated in digest JSON.
• Compatibility boundary: existing runtimeEventsPath behavior remains unchanged; reward-hack scanning still uses model-visible runtime events, not the diagnostic trace ledger.

Validation

• npm run -w @maka/headless test — 405 tests pass
• npm run -w @maka/headless typecheck
• git diff --check

Follow-ups (not in this PR)

• A/B harness to compare a hand-authored / opencode-derived baseline prompt against the current seed using the existing acceptance policy.
• Real unattended RSI run with this signal enabled, after review, to see whether coverage-regression discards decrease.

Part of #64.