fix(poteto-mode): bound delegation and carry the checkpoint below the floor - #16
Conversation
A loop round-one session followed every gate yet died on its own deadline: its scout read an 800KB corpus for an inventory and the main session set a ten-minute blind wait against an eight-minute budget. Pins the depth-bound and wait-sizing clauses before the prose lands.
The delegation defaults make depth part of the dispatch contract and size wait timeouts to the work still owed, preferring a short wait plus a jobs re-check over one long blind block. Records loop round one in CHANGES.md.
Round two's docs session dropped the one-line checkpoint the other six below-floor sessions emitted unprompted; the trigger's phrasing left it optional below the floor. Pins the universal clause before the prose lands.
Below-floor replies now carry one n/a checkpoint line so an audit can tell poteto-mode ran at all. Records loop round two in CHANGES.md.
daf05bc to
6dd91ad
Compare
Independent verdict: FAILPR head Local gates (all PASS)
Note (non-blocking): the PR body says Live gate: probe PASS, sessions 3/4 — one (a) run FAILOverride mechanism: candidate copied to Prompt (a) — scout inventory, run 1: FAIL (Deadline-exceeded, no final reply). Prompt (a) — scout inventory, run 2: PASS. Prompt (b) — hypothetical docs draft, run 1: PASS. 98.95s, reply ends Why FAILThe acceptance contract for this verification makes any (a) run that misses the 420s budget a FAIL. Run 1 of (a) missed it: deadline-exceeded with no final reply. The two new SKILL.md clauses demonstrably shape behavior (both runs dispatched depth-bounded scouts and used short Machine restored
|
The retry verifier's failing run followed both new clauses yet its dispatch still told the scout to read each test file, and the scout spent five and a half minutes on a twelve-suite inventory before the 420s wall. Pins the labels-not-bodies clause before the prose lands.
The delegation defaults name the unit of inventory work: describe and it labels, exports, headings, directory shape, with file bodies opened only as the range that answers a specific doubt. Records verifier round two in CHANGES.md.
Verdict: FAIL — head d1098a1 (base 7f9211d)Third verification attempt (verifier 16c). All local gates pass and three of four live sessions clear, but run 2 of the delegation scenario missed the scout-transcript size gate: 437,928 bytes against the 400KB cap. Its dispatch omitted the labels-not-bodies depth clause, the scout read twelve full test-file bodies, and the transcript blew the cap even though the session itself finished inside budget. The new sentence is pinned and one run honored it verbatim; the second run shows it does not yet reliably shape the dispatch. Gates
Probe note: the prescribed symlink target Live evidence (each prompt run twice,
|
Verdict: PASS+NOTES — head d1098a1 (base 7f9211d)Fourth verification attempt (verifier 16d), under a corrected acceptance contract that gates user-facing behavior only. All local gates pass, the probe returns READY=yes, all four live sessions clear every hard gate, and the machine is restored. Depth-clause fidelity, scout transcript size, and hub-wait shape are reported below as evidence, not gates. Gates
Live evidence (each prompt run twice,
|
1 participant
Stacks on #15, extends issue #14.
Why
A
/loopaudit battery of realomp -psessions ran against the candidate build until a full round came back clean. Each round launched seven varied poteto-mode scenarios in a fresh worktree from main with #15 merged in, audited every transcript from the skill-load dispatch onward, and fixed what the evidence justified. Round one found a session that followed every gate yet died on its own deadline. Round two found a below-floor session silently dropping the checkpoint line. Round three came back clean on all seven.Scope
SKILL.mddelegation defaults bound two things. Depth is part of the contract: an inventory or summary assignment says what depth it needs and forbids full-corpus reads the answer does not require. Ahub waittimeout is sized to the work still owed after the result, with a short wait plus ahub jobsre-check preferred over one long blind block.SKILL.mdcheckpoint trigger now covers below-floor replies: one line,throughput checkpoint: n/a, <kind>. The checkpoint is how an audit knows poteto-mode ran at all.check-playbooks.test.tspins both rounds red-first, two new describes.CHANGES.mdrecords loop rounds one and two under the third-audit section.Tradeoffs
The below-floor checkpoint line adds one sentence to every small reply. Six of seven below-floor sessions already emitted it unprompted, so the clause makes existing behavior uniform rather than adding ceremony.
Blast Radius
Sessions that delegate read-only research read two tighter sentences and size their waits. The scripts suite grows from 178 to 182 tests. Nothing migrates and no surface changes shape.
Verification
Local gates on the branch tip.
bun run testinplugins/pstack/skills/poteto-mode/scriptsreports 180 pass and 0 fail.bun run typecheckruns all three strict projects clean.bun test tests/reports 18 pass.bash tests/skill-collision-repro.shreports all five invariants ok at 2.1.5.Live evidence:
pstack-loop-candidateon a copy of this worktree, releasepstack@omp-pstack2.1.4 disabled during the batteries. Version probes repliedREADY=yesfor each new clause before its battery ran. The machine was restored after;omp plugin listservespstack@omp-pstack (2.1.4).omp -psessions, three full rounds of seven scenarios each in the worktree.skill://poteto-modeload onward for todolist-first, marks-as-landed, playbook steps with visible skips, the checkpoint line, how routing or skip markers, edit discipline, and reply format.hub wait.throughput checkpoint: n/a, read-only drafting.with a substantive how-skip reason. The delegation session finished in budget. Zerolevel:errorlines across the six battery process logs.Verifier rounds
The first independent verifier for this PR returned FAIL on a doubled live gate: one delegation run followed both new clauses yet still hit the 420s wall, because its dispatch told the scout to read each of the twelve test files and the scout spent five and a half minutes on bodies. The retry verifier's evidence also forced the parent PR's reply-contract fix. This PR now also grounds inventory delegation in names and structure over file bodies, pinned red-first. Verdict comments: #16 (comment) and the PR 15 thread.