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
This is the actual autonomous LOOP that Phase 3 exists to build: given a plan/acceptance-criteria packet (from Phase 2, out of scope here) and a resolved CodingAgentDriver, repeatedly invoke the driver, self-review the resulting diff against src/rules/predicted-gate.ts (the byte-identical predicted-gate target named as THE self-review target for an autonomous miner) plus the acceptance-criteria file, and DECIDE — autonomously, with no human in the loop at this stage — whether to keep iterating, stop and hand off to Phase 4 submission, or abandon the attempt. This decision logic is the precise chokepoint the fixed architecture skeleton's safety-tier system reserves for the owner: it is the trigger surface for "does the system keep trying, or does it eventually open a PR" without a human approving each step.
Tagged maintainer per this phase's explicit safety split: the driver seam, drivers, and harness primitives (all prior issues in this batch) are reviewable, no-autonomous-trigger contributor work; this orchestration control flow is not — it is adjacent to the #1 slop-at-scale strategic risk (an autonomous fleet maximizing gate-pass rate can mass-produce gate-passing-but-low-value PRs) and the owner has stated intent to build/review this personally.
Deliverables
packages/gittensory-engine/src/miner/iterate-loop.ts implementing the create->score->self-review->decide control flow, consuming a CodingAgentDriver + the predicted-gate self-review target
A bounded max-iteration/max-cost ceiling enforced INSIDE the loop itself (not left to an external caller to remember to enforce)
Every iteration's decision (continue / stop-success / abandon) is recorded via the attempt-log primitive (from the wave-2 batch) so the decision trail is auditable after the fact
Default behavior when the loop cannot conclusively determine pass/fail (e.g. self-review itself errors) is to ABANDON, never to optimistically hand off to submission — downgrade-only on ambiguity
Owner-reviewed test coverage proving the loop never hands off to Phase 4 submission on a predicted-gate FAIL or an ambiguous self-review result
References
/Users/shadowbook/Documents/gittensory/.claude/worktrees/quizzical-engelbart-34f3be/src/rules/predicted-gate.ts (263 lines) — the byte-identical self-review target this loop must consult every iteration; see especially buildPredictedGateVerdict's conclusion/blockers fields
packages/gittensory-engine/src/miner/coding-agent-driver.ts, attempt-log.ts, acceptance-criteria.ts (all wave-2 primitives this loop composes, prior issues in this batch)
User memory note "gittensory-miner-autonomy-roadmap": "the feat(docs): add install site and mcp diagnostics #1 slop-at-scale risk" and "never auto-submit (P4) before governor+caps (P5)" ordering rule — this issue is exactly that boundary, held to maintainer per the phase brief's explicit instruction
This is the actual autonomous LOOP that Phase 3 exists to build: given a plan/acceptance-criteria packet (from Phase 2, out of scope here) and a resolved
CodingAgentDriver, repeatedly invoke the driver, self-review the resulting diff againstsrc/rules/predicted-gate.ts(the byte-identical predicted-gate target named as THE self-review target for an autonomous miner) plus the acceptance-criteria file, and DECIDE — autonomously, with no human in the loop at this stage — whether to keep iterating, stop and hand off to Phase 4 submission, or abandon the attempt. This decision logic is the precise chokepoint the fixed architecture skeleton's safety-tier system reserves for the owner: it is the trigger surface for "does the system keep trying, or does it eventually open a PR" without a human approving each step.Tagged
maintainerper this phase's explicit safety split: the driver seam, drivers, and harness primitives (all prior issues in this batch) are reviewable, no-autonomous-trigger contributor work; this orchestration control flow is not — it is adjacent to the #1 slop-at-scale strategic risk (an autonomous fleet maximizing gate-pass rate can mass-produce gate-passing-but-low-value PRs) and the owner has stated intent to build/review this personally.Deliverables
packages/gittensory-engine/src/miner/iterate-loop.tsimplementing the create->score->self-review->decide control flow, consuming aCodingAgentDriver+ the predicted-gate self-review targetReferences
/Users/shadowbook/Documents/gittensory/.claude/worktrees/quizzical-engelbart-34f3be/src/rules/predicted-gate.ts(263 lines) — the byte-identical self-review target this loop must consult every iteration; see especiallybuildPredictedGateVerdict'sconclusion/blockersfieldspackages/gittensory-engine/src/miner/coding-agent-driver.ts,attempt-log.ts,acceptance-criteria.ts(all wave-2 primitives this loop composes, prior issues in this batch)maintainerper the phase brief's explicit instruction