Force overrides only cleanliness. Bind approval to Git HEAD/tree, index and dirty/untracked content fingerprints; recheck at CompleteGoal. Require investigative read-only supervision instead of accepting gate errors as experiment blockers.
Use explicit hello requests and replies, keep workers unready until model restoration succeeds, distinguish paused peers, and align all goal readers at the Log boundary. Add two-real-adapter handshake and failed-Ready/clear-during-wait regressions; warn once for unavailable supervisor context usage.
Restore bindings before model availability checks, require explicit reconnect/restart recovery, detach inactive plans, remove the general Intercom actuator, and reject placeholder evidence without hashing the plan log. Preserve synchronous handoff-before-ack; document that the void SDK cannot confirm durable message delivery.
Keep choices separate by role, ignore automatic restores, and fail when a remembered model is unavailable. Restore the worker model only after the planning fork is ready.
Borrow the provider tracker queries from cecb1e9. Keep unavailable state unknown, bind incremental views to acknowledged source entries, reset after compaction, bound serialized payloads, and recheck tracked work at sign-off.
Scope messages to each pairing, wait for supervisor readiness, retain instructions in session history until acknowledgment, and rejoin after disconnect. Reuse installed Intercom or load the package dependency. Validate over an isolated real broker; existing user panes are untouched.
Render exact instructions, restore monitoring/read-only tools on resume, report actual Pi idle state, reject stale approval views, and stop completed-plan timers. Ask for brief evidence-based judgment. Add real TUI rendering and lifecycle regressions; retain explicit limits on background state and unmeasured cost benefit.
Co-Authored-By: Pi/OpenAI <288921227+claudypoo@users.noreply.github.com>
"Ready?" over an unread file is not a review -- the only copy of the plan was
inside a collapsed edit tool call. Print the working set before the menu.
The 4th option compacts the planning conversation before the work turn starts.
session_compact already re-sends the whole plan file, so the exploration is
summarized away and the agreed goals are not.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
The judge has been read-only with no bash since the rewrite, but the README still said it runs the
goal's verify command. That is the exact thing a dogfooding agent got wrong out loud.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
Re-running verify was fine for 'npm test' but a footgun for ML workflows where verify may be a
10-hour training run -- and bash made 'read-only' nominal anyway (it could mutate). The agent now
runs verify itself and saves the output as evidence. The judge checks, in order: anything here /
quoted+attributed / provenance / quotes match disk / substance. Matches the cooperative-but-
confused threat model: reading real artifacts catches confusion; execution only defended against
deliberate forgery, which is out of scope.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
'Did the judge really re-run verify?' was unanswerable post-hoc; now every sign-off's full
judge output survives on disk.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
Dogfood exit interview: agent bookkeeping is the drift point (tick after accept was the
step most likely forgotten), and 'committed artifact' language implied the judge sees HEAD.
Tick is exact-subject match via the existing GOAL_LINE regex; on drift the result explicitly
asks the agent to tick, so neither path is silent.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
CompleteGoal.execute now delegates to a pure decideSignOff(input, signal,
runJudgeFn) that takes an injected judge runner. judgeModel is never checked
pre-emptively: null just omits --model (buildJudgeArgs), so pi's configured
default runs the judge. The only producers of accepted_inconclusive are now the
judge-error and no-VERDICT paths -- i.e. 'the judge ran but failed', never 'no
model'. Wording (log/result/docstring/README) updated to say 'ran but failed'.
Adds test/decide-signoff.test.ts: null judgeModel still reaches runJudge (no
pre-emptive return); judge error/timeout/no-VERDICT -> accepted_inconclusive
with a 'ran but failed' reason. 11 tests pass, typecheck + lint clean.
The v1 lesson: the parser existed so TypeScript could read goals.md, but every
reader is a model. v2 injects .pi/plan.md verbatim each turn, teaches the format
as a convention, and hands the whole file to the judge, which now does the goal
matching (tolerates wording drift), evidence validation, verify execution, and
format reading that v1 did in code. 1874 -> 530 lines.
Deleted: plan-file.ts + tests, JSON-stream judge transport, custom tool
rendering, review menu + $EDITOR + newSession dance, pruneCompleted, unwired
continuation/loopJudge prompts, MUTATING_BASH_PATTERNS. CompleteGoal's only
write is the ## Log sign-off line (the audit trail); the agent ticks [x] itself.
Judge runs with --no-extensions so a broken global extension can't take down
sign-offs (pi-hermes-memory currently does exactly that). File renamed
goals.md -> plan.md.
UAT (real judge subprocess on a toy repo, /tmp/claude-goals-uat/judge-*.log):
accept with verify run + byte-check, reject on placeholder evidence + missing
file, accept under drifted goal wording.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
wassname