feat(method): six-phase-loop — 8→6 phase merge, bundle-disclosed guides, persona presets, build-entry spec echo by pilotspacex-byte · Pull Request #148 · pilotspace/ADD

pilotspacex-byte · 2026-07-14T02:44:11Z

six-phase-loop — TRUE 8→6 phase merge + guides disclosed into bundle subagents

Stacked on #147 (feat/call-floor). Born from the operator's asks: "compress steps 1-2 and 6-7" · "make expert subagents for each step" · "disclose step skills into the subagents, not the main agent". All four design decisions were interview-confirmed before drafting (recorded in MILESTONE.md Ground).

Close review: 6/6 tasks gate PASS · 7/8 exit criteria evidenced with (verify:) citations · HELD OPEN on the paid WM1 re-measure (human-gated). Full fence 3554/3554 OK, zero tests weakened; suite grew 3508→3554.

Every commit (pre-squash census)

commit	task	what it does
`37d5a80`	—	milestone draft (await-confirm)
`fa12a4d`	1/6 phase-merge-specify	scenarios→specify; §-sections never renumber; NEW `_PHASE_SECTIONS` explicit table kills the phase-index-vs-§-section off-by-one class
`8467b46`	2/6 phase-merge-verify	observe→verify (final list `specify·plan·tests·build·verify·done`); skip grammar RETIRED whole (`_SKIPPABLE_PHASES=()`, vestigial `skips:` noted loud at gate, never a die); legacy phase tokens normalize-on-read; `--fill`/`--section` repaired onto the table
`7ffdd2d`	3/6 guide-recut	6 guide files (1-specify absorbs 2-scenarios; 6-verify absorbs 7-observe as the post-gate Observe block); phases pool INVERSE-rebaselined 41605→40205 (net-won ÷ ratio, signed)
`e3e0164`	4/6 bundle-disclosure	delegating: each roster agent loads its OWN bundle guides at spawn, the orchestrator reads only SKILL.md (9442B < 9500 ceiling); inline lane stays first-class
`be74a90`	5/6 persona-presets	per-phase teacher-grade `Preset (<phase>):` stances replace the thin generic fallback; project-persona routing stays first; never lowers a gate
`514df5f`	6/6 build-entry-spec-echo	the tests→build tick prints §1 Must/Reject + the frozen §3 contract head (`_spec_echo`, pure read, fail-open tail of `_build_entry`, both entries) — self-proved live on its own re-cross
`16bbf27`	—	close-review bookkeeping (criteria evidenced)

Pins & floors

ENGINE_MD5 → bf89b033 (PKG pin correctly NOT re-aimed — add_engine/ digest unchanged at guide/agent tasks touched no engine) · SEAMS _declared_scope pin → :5653 · 3-tree skill parity + agents canonical→_bundled→.claude parity test-pinned · gate semantics byte-identical (one recorded outcome, security HARD-STOP, tamper/scope guards untouched).

What is proven vs pending

Proven mechanically (in this PR's tests): 6-phase lifecycle end-to-end, loud-safe legacy migration, guide census + byte ceilings, disclosure split, preset census, spec-echo behavior + fail-open.
Pending (holds the milestone open): paid WM1 re-measure — fidelity ≥ 0.97 held and calls ≤ 12 on the live benchmark. Run or accept-partial before merge.

LESSON recorded: declare §4 tests as the exact FILE — a dir-form declaration sweeps build targets into the tamper snapshot (recovered via signed re-cross, the expectations-first recipe).

CI fresh-checkout failure on PR #148: test_engine_twins_carry_the_echo asserted parity on the two gitignored dogfood twins (.add/tooling/ and add-method/.add/tooling/), which do not exist on a fresh clone — the known gitignored-twins class (build-strategy-facets lesson). The tracked _bundled twin now binds unconditionally; the dogfood twins are checked only when present (fresh-checkout skip-tolerance precedent ba09498). No assertion weakened where the files exist. refs: six-phase-loop milestone, PR #148 author: Tin Dang <tindang.ht97@gmail.com>

8->6 phase merge (specify absorbs scenarios; verify absorbs observe) + phase-guide disclosure into the bundle-owning roster agents so the orchestrator loads only SKILL.md + per-phase persona presets + the build-entry spec echo. Human decisions from the 2026-07-14 interview recorded in Ground; tasks gated until milestone-confirm. refs: milestone six-phase-loop (await-confirm) author: Tin Dang <tindang.ht97@gmail.com>

…x-phase-loop 1/6) TRUE phase merge, first of two (human-decided over recipe-ticks): the scenarios PHASE folds into specify — one drafting phase now produces §1 rules AND §2 Given/When/Then. The TASK.md §-section shape is UNTOUCHED (sections are the stable API; only the lifecycle list shrinks). Engine (synced x3 — tooling · .add/tooling · _bundled): - constants.py: PHASES=(specify,plan,tests,build,verify,observe,done); _SKIPPABLE_PHASES=(observe,); PHASE_GUIDE specify action gains the Given/When/Then duty; PHASE_OWNER/PHASE_AGENT drop scenarios; PHASE_GROUPS DIRECTION=(specify,plan,tests) - add.py: _normalize_phase_tokens legacy map gains scenarios->specify (silent, TOTAL, idempotent, normalize-on-read — expectations-first precedent); _FRONT_PHASES/(spawn-hints/guide-file map/freeze early-phase check/skip-rationale regex follow); _RETIRED_SKIP_TOKENS: a pre-merge `skips: scenarios` header is tolerated LOUD (note at the observe crossing, filtered from the skip set) while non-skippable tokens keep the die — the frozen M13 pin (zero skip logic on non-skippable crossings) holds; task_phases: explicit _PHASE_SECTIONS table replaces the index-derived n->section mapping (the named off-by-one trap, hit live) - installer showcase twins (bin/cli.js BRAND_LOOP · _installer.py _LOOP): 6-step loop, grounded by test_onboarding_brand Tests: NEW test_phase_merge_specify.py (8 red->green: phase list + maps, advance lands plan, scenarios refused, legacy state normalizes, old skip declaration tolerated loud, bad token still dies). ~30 fixture suites repaired as the declared sequence-shift ripple (test_min_pillar walk now 4 advances to verify; test_add advance pin; onboarding showcase slice; report dots; phase-detail blocks; fast-lane skips rewritten to observe) — zero tests weakened. Full fence 3517/3517 OK. Pins: ENGINE_MD5 -> b27ce845 · ENGINE_PKG_MD5 -> 870a4ce0 · SEAMS _declared_scope anchor -> add.py:5631. Gate: PASS (compound tick from build; scope widened +installer twins via sanctioned re-cross). refs: six-phase-loop milestone (37d5a80), stacked on feat/call-floor (#147) author: Tin Dang <tindang.ht97@gmail.com>

…ase-loop 2/6) The second TRUE phase merge: the observe PHASE folds into verify — verify owns evidence + gate AND the watch/spec-delta duty. §7 OBSERVE stays as a section (the stable API) rendered under the verify block, exactly as §2 renders under specify. The lifecycle is now specify · plan · tests · build · verify (+ done). Engine (synced x3): - constants.py: PHASES=(specify,plan,tests,build,verify,done); _SKIPPABLE_PHASES=() — nothing is skippable; PHASE_GUIDE verify action gains the watch+spec-delta duty; PHASE_GROUPS VERIFY=(verify,); PHASE_OWNER/PHASE_AGENT drop observe - add.py: _legacy gains observe->verify (normalize-on-read; a recorded gate survives); _RETIRED_SKIP_TOKENS={scenarios,observe}; the cmd_advance skip block PRUNED whole (dead once the tuple emptied — the old M13 zero-skip- logic placement is now universal); gate/completion is the ONE seam that reads a vestigial `skips:` header (one loud note, never a die) and the new home of the fold nudge; advance --fill + status --section resolve through _PHASE_SECTIONS (verify: (6,7)) — repairing the two ordinal->section mappings that broke silently at phase-merge-specify (plan filled §2, not §3); membership lists/spawn hints/guide map/help text swept - templates: the fast template's skips scaffold removed (grammar retired end-to-end); both templates' phase-marker comments show the 6-token flow - installer showcase twins: BRAND_LOOP/_LOOP = 5 pre-done steps Tests: NEW test_phase_merge_verify.py (13: list+maps, advance/gate land done with the fold nudge, observe refused, legacy state normalizes keeping its gate, vestigial skips noted at gate never die, _PHASE_SECTIONS + --fill/ --section table pins). Fence r1 ripple: 49 failures/17 suites repaired — test_fast_lane_skips REWRITTEN to pin the retirement (resolver keeps naming bad tokens; historic recorded skips keep status/audit surface); fold nudge test re-aimed to gate; report dots 6-wide; phase-detail 5 blocks with §7 under verify; onboarding showcase 5 steps; len/range pins swept. Zero tests weakened. Full fence 3527/3527 OK. Pins: ENGINE_MD5 -> 5d5e0538 · ENGINE_PKG_MD5 -> fc40ad47 · SEAMS _declared_scope anchor -> add.py:5603. Gate: PASS (compound tick from build; the new fold nudge printed on this task's own gate — self-proved). refs: six-phase-loop milestone, stacked branch feat/six-phase-loop author: Tin Dang <tindang.ht97@gmail.com>

…six-phase-loop 3/6) The skill's guide tree follows the engine's merged lifecycle: 1-specify.md absorbs 2-scenarios' load-bearing duties (gherkin Given/When/Then format, one scenario per Must AND per Reject, the And-unchanged clause on every rejection, the edge-case sweep, the exit-gate boxes); 6-verify.md absorbs 7-observe's as a post-gate Observe block (scope-of-impact release, scenarios-as-monitors, the next SPEC delta, the voice-delta -> soul.md route, the persona tag line, the self-improve.md map pointer). The two absorbed files are DELETED from all three guide trees. SKILL.md's phase table shows the 6-row loop (specify row owns §1+§2, verify row §6+§7); 9355 B, under the orient-split 9500 ceiling. soul.md re-routes its Emit line to the verify guide. Pools: phases baseline re-anchored 41605 -> 40205 — the INVERSE of the rebaseline-for-new-surface method: deletion is won ground the ceiling locks in, so the baseline shrinks by NET won bytes ÷ ratio (deleted 3855 B minus the absorbing guides' +2735 B = 1120 net; −⌈1120/0.80⌉=1400), ratio 0.80 kept exactly. test_ground_anchor_sha's baseline floor forward-migrated 40280 -> 40205 with the accounting signed. Tests: NEW test_guide_recut.py (8: deletion ×3 trees · duty markers in both absorbing guides · no live prose names a deleted file · table rows/§-spans · ceiling held). Ripple: 11 failures/10 suites re-aimed (wording-surface census 32→30 in two pins; soul/persona/ADR/xml-census/per-step-hook pins moved onto the absorbing guides; pool lists trimmed). Zero tests weakened. Full fence 3535/3535 OK. Engine untouched — no MD5 re-aim. refs: six-phase-loop milestone, stacked branch feat/six-phase-loop author: Tin Dang <tindang.ht97@gmail.com>

… agents (six-phase-loop 4/6) The user's directive made mechanism: when the orchestrator delegates a phase, the roster bundle agent loads its OWN phase guides at spawn — the orchestrator reads only SKILL.md. The inline lane stays first-class (load the one phase guide yourself, unchanged). - add-design.md: NEW "Load your bundle guides" section (phases/0-setup.md on fresh projects + phases/1-specify.md + phases/3-plan.md); span prose re-cut to setup · specify · plan (grounding/scenarios/contract live inside them); description spawn steps -> SETUP, SPECIFY, or PLAN; Return enum follows - add-build.md: loads phases/4-tests.md + phases/5-build.md at spawn - add-verify.md: loads phases/6-verify.md (carries the post-gate Observe block); owns verify end-to-end — OBSERVE is no longer a spawn step; Return enum verify-only - SKILL.md flow lead-in states the split in one sentence: inline -> load only the phase you are in; delegating -> the Bundle-column agent loads its own guides, you read ONLY this file. 9442 B, under the orient-split 9500 ceiling. Synced x3 skill trees; agents synced canonical -> _bundled + .claude installed copies. Tests: NEW test_bundle_disclosure.py (6: per-agent guide census · no deleted guide or retired step name — path-qualified, docs chapters keep their names · disclosure sentence · ceiling · x2 parity). Full fence 3541/3541 OK. Engine untouched — no MD5 re-aim. refs: six-phase-loop milestone, stacked branch feat/six-phase-loop author: Tin Dang <tindang.ht97@gmail.com>

…ces (six-phase-loop 5/6) The interview decision made mechanism: per-phase expertise via PRESETS on the lean 5-agent roster, not a per-step agent rebuild. Each bundle agent's thin "use a generic engineer" fallback upgrades to named expert stances per owned phase, distilled from the vendored teacher library: - add-design.md: Preset (specify) — a 15-year domain analyst who asks rather than assumes, every Reject earns a named error code (setup rides this); Preset (plan) — a systems architect, contract-first, cites only anchors that exist, sizes the change's reach before freezing scope - add-build.md: Preset (tests) — red for the RIGHT reason, one test per scenario, behavior never internals; Preset (build) — the smallest honest diff in small reviewable batches, a test is never touched to go green - add-verify.md: Preset (verify) — adversarial reviewer then reliability analyst, refutes the green before trusting it, security always HARD-STOP, spans the post-gate Observe duties (§7) Presets are the FALLBACK tier only: project-persona routing stays first (the flow: selection sentence untouched); a preset never blocks and never lowers a gate. Synced canonical -> _bundled + .claude installed copies. Tests: NEW test_persona_presets.py (4: per-phase preset census · no bare generic fallback left · routing-first + never-lowers-a-gate survive · x2 parity). Full fence 3545/3545 OK. Engine untouched. refs: six-phase-loop milestone, stacked branch feat/six-phase-loop author: Tin Dang <tindang.ht97@gmail.com>

…WHAT to build (six-phase-loop 6/6) The interview decision made mechanism: at every successful tests->build tick the engine prints `build to (frozen plan):` + the §1 Must/Reject bullet first-lines + the frozen §3 contract-fence head, so the builder starts from the spec on the screen, not from memory. - add.py: NEW _spec_echo (PURE read beside _scope_echo; bullet walk stops at the next top-level `Key:` line, continuations never echo) + a wrapped tail call in _build_entry — AFTER every guard and snapshot, so a refused entry never echoes and BOTH entries (advance + the `phase build` override) fire; fail-open: exit code, state.json, and TASK.md are identical either way - ENGINE_MD5 re-aimed bf89b033; ENGINE_PKG_MD5 deliberately NOT re-aimed (add_engine/ untouched — the digest proved it unchanged); SEAMS _declared_scope pin 5603->5653; engine synced x3 twins - SELF-PROVED: the task's own re-cross into build rendered its Must/Reject/ contract echo live Tests: NEW test_build_entry_spec_echo.py (9: echo at advance + at the phase override · line order · forced-raise never blocks the tick · malformed §1/§3 partial-echo no-crash · refused crossing never echoes · --skip-freeze DRAFT crossing still echoes · pure-read byte check · engine-twin parity). Full fence 3554/3554 OK. LESSON: declare §4 tests as the exact FILE — a dir-form declaration sweeps the build targets into the tamper snapshot (recovered via signed re-cross, the expectations-first recipe). refs: six-phase-loop milestone, stacked branch feat/six-phase-loop author: Tin Dang <tindang.ht97@gmail.com>

…t criteria evidenced Milestone close-review bookkeeping: all six task boxes checked, seven exit criteria checked with (verify: ...) citations (test suites + fence 3554/3554 + the self-proved build-entry echo), rationale placeholder filled with the confirmed intake classification. The goal-gate holds the milestone OPEN on the eighth criterion — the paid WM1 re-measure (fid >= 0.97, calls <= 12) — which is human-gated: run or accept-partial. refs: six-phase-loop milestone, stacked branch feat/six-phase-loop author: Tin Dang <tindang.ht97@gmail.com>

CI fresh-checkout failure on PR #148: test_engine_twins_carry_the_echo asserted parity on the two gitignored dogfood twins (.add/tooling/ and add-method/.add/tooling/), which do not exist on a fresh clone — the known gitignored-twins class (build-strategy-facets lesson). The tracked _bundled twin now binds unconditionally; the dogfood twins are checked only when present (fresh-checkout skip-tolerance precedent ba09498). No assertion weakened where the files exist. refs: six-phase-loop milestone, PR #148 author: Tin Dang <tindang.ht97@gmail.com>

+    # echoes; fail-open, so exit code, state, and TASK.md are identical either way.
+    try:
+        _spec_echo(root, slug)
+    except Exception:


+    # echoes; fail-open, so exit code, state, and TASK.md are identical either way.
+    try:
+        _spec_echo(root, slug)
+    except Exception:


…als follow-on draft (#150) Human decisions 2026-07-14: run the paid WM1 re-measure, then "Close + draft follow-on milestone". - WM1 re-measure (3 reps, pinned sonnet meter, engine = merged main 8360b06): fidelity 0.98 x3 / 0 regressions / oracle 1.0 -> criterion MET; calls 20.7 mean vs <=12 -> UNMET, HUMAN-ACCEPTED-PARTIAL at close; cost $3.17 mean = cheapest honest ADD round (-10% vs ceremony). Full record + waste anatomy: benchmark/results/2026-07-sixphase-remeasure.md (runs archived benchmark/runs/sixphase-r{1,2,3}). - six-phase-loop -> done (RETRO.md written; 21 open deltas left for the human fold review). - NEW milestone draft call-residuals (--await-confirm, tasks gated): the four message-layer levers the anatomy named — init-idempotent-nudge · scope-first-draft · status-orientation-diet · help-habit-kill — targeting the ~11-12 call landing with zero enforcement-path changes. refs: six-phase-loop close, PRs #146/#149/#148 merged author: Tin Dang <tindang.ht97@gmail.com> Co-authored-by: Tin Dang <tindang.ht97@gmail.com>

TinDang97 force-pushed the feat/call-floor branch from 9ed0412 to e35c85d Compare July 14, 2026 03:41

pilotspacex-byte changed the base branch from feat/call-floor to main July 14, 2026 03:44

TinDang97 added 9 commits July 14, 2026 10:44

TinDang97 force-pushed the feat/six-phase-loop branch from bb80287 to 8360b06 Compare July 14, 2026 03:44

github-code-quality Bot found potential problems Jul 14, 2026

View reviewed changes

pilotspacex-byte merged commit 0dbc082 into main Jul 14, 2026
8 of 9 checks passed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

feat(method): six-phase-loop — 8→6 phase merge, bundle-disclosed guides, persona presets, build-entry spec echo#148

feat(method): six-phase-loop — 8→6 phase merge, bundle-disclosed guides, persona presets, build-entry spec echo#148
pilotspacex-byte merged 9 commits into
mainfrom
feat/six-phase-loop

pilotspacex-byte commented Jul 14, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

pilotspacex-byte commented Jul 14, 2026

six-phase-loop — TRUE 8→6 phase merge + guides disclosed into bundle subagents

Every commit (pre-squash census)

Pins & floors

What is proven vs pending

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants