You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Commit 54e93b7
Browse filesBrowse the repository at this point in the historyBrowse files
fix(harness): sixth-pass — correct four defects the bots found in the fifth pass
The review bots re-reviewed the previous commit and caught real problems in it.
Four were regressions I introduced; two were weaknesses in how I fixed things.
Regressions in the fifth pass, now corrected:
- The per-result dispatch cap was extended to FAILED tool results, but
`ToolResultCap.apply` appends a hint that says *"The tool call succeeded"*.
An oversized error was therefore persisted with success wording, so the model
could read a real failure as a truncated success. `apply()` now takes an
`outcome` and emits an error-specific hint that states the failure is real.
- The nudge arbiter was changed to replace pending directives by SOURCE alone.
That is wrong: four independent detectors register under
`starvation_breaker` (`doom_loop_nudge`, `doom_loop_status_check`,
`repeat_signature`, `starvation`) and they fire at different points in a step,
so a later WEAKER detector could clobber a stronger directive that had already
fired — the opposite of the intended fix. Replacement is back to source+kind,
and `take()` now ranks kinds by explicit strength within the winning source.
- The prompt-retry acceptance probe failed OPEN: if the probe itself threw, the
message was treated as absent and the task was resent — exactly the
duplicate execution the mechanism exists to prevent. Acceptance is now
three-valued. Only a definitive 404 from a reachable server permits a retry;
an unknown state fails the run with a clear message rather than risking a
second execution or a silent hang.
- Stripping `ALTIMATE_RUN_MODE` from bash child environments also removed an
explicit opt-out (`=0`/`false`), so a nested `run` would re-apply the default
and turn run mode back ON. Only an ACTIVE marker is stripped now.
Weaknesses in the fifth pass, now addressed:
- `ToolResultCap.resolve` resolved the safety fraction AFTER the unknown-model
branch returned, so a configured fraction never scaled that fallback.
- The run-mode child-env test asserted on `bash.ts` source text, which cannot
catch the strip being relocated, applied to a different object, or skipped —
the same false-confidence problem flagged elsewhere in this review. The logic
is now the exported `stripRunModeMarkers()` and the test exercises it.
Tests: strongest-wins and weaker-does-not-displace ordering for every
`starvation_breaker` kind; error vs success truncation wording (and that the
error hint still respects the cap); configured fraction scaling the
unknown-model fallback; and behavioural coverage of the marker strip including
opt-out survival.
Gates: typecheck clean, marker check clean, 2411 pass / 0 fail.
Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01VqnuBDGkh1ZT65Ti7e6DHZ
// altimate_change start — outcome-accurate hint (see `opts.outcome`).
102
120
consthint=
103
-
"The tool call succeeded but the output exceeded the per-result context budget and was truncated before dispatch. Re-run the tool with a narrower query (filters, LIMIT, offset/limit) to view specific sections."
121
+
opts?.outcome==="error"
122
+
? "The tool call FAILED and its error output exceeded the per-result context budget, so the error text below was truncated before dispatch. The failure is real — do not treat this as a successful result. Re-run with a narrower scope if you need the full error."
123
+
: "The tool call succeeded but the output exceeded the per-result context budget and was truncated before dispatch. Re-run the tool with a narrower query (filters, LIMIT, offset/limit) to view specific sections."
0 commit comments