feat: science-project memory — a fresh chat resumes a project by default

[Jammy2211]

Overview

Point a fresh agent chat at a science project folder and it should work out where the project got to, how to resume, and what its history is — by default, without being asked "what have we done?". Today it does not: the generated project AGENTS.md has no session-start section, nothing reads the journal on arrival, results_summary.md is consumed at Publish but never templated, results/manifests/ is aspirational (empty after a full real campaign), and profile.md conflates user-level facts with the per-project science goal.

This task makes that true for projects scaffolded by the assistants' start-new-project skill, and lands the change once — via a new sync mode on the PyAutoBrain clone conductor that re-applies the reference assistant's _SHARED_GENERIC diff to its siblings as a reviewable patch, rather than four hand edits.

euclid_assistant is a different design and is explicitly out of scope.

Plan

• Phase 1 — PyAutoBrain. Add a sync mode to the clone conductor: given a reference assistant and a commit range, take the reference's diff to the _SHARED_GENERIC files and apply it to each sibling as a three-way patch, reporting per-file/per-sibling applied cleanly / hunks rejected / file absent. Dry-run by default; --apply writes. Domain-adapted lines in siblings survive; conflicts surface for a human instead of being resolved silently. Wire into bin/pyauto-brain clone sync, document in the conductor's AGENTS.md, unit-test with a temp-dir fixture.
• Phase 2 — autolens_assistant (the reference copy). Deliverables A–D: a session-start block in the generated project AGENTS.md (+ a project CLAUDE.md that @-imports it); a state.md head pointer (templated, scaffolded, rewritten-not-appended each session, wired into the end-of-session journal ritual); a templated results_summary.md with a covers_through: stamp; results/manifests/ deleted from the scaffold and from every sentence that promises it; and profile.md split at the user/project seam.
• Phase 3 — propagate. Run the Phase-1 sync dry-run from autolens_assistant to autocti_assistant, autogalaxy_assistant, autofit_assistant. Apply and PR the two that take the patch cleanly; autofit_assistant (diverged by ~343 lines) is not applied — its dry-run report is posted here for a human.

Detailed implementation plan

Work Classification

Library (three independent repos; no workspace follow-up — the assistants are their own workspaces).

Affected Repositories

• PyAutoBrain (primary)
• autolens_assistant
• autocti_assistant
• autogalaxy_assistant
• autofit_assistant (dry-run only — no edits this task)

Branch Survey

Repository	Current Branch	Dirty?
./PyAutoBrain	main	clean
./autolens_assistant	main	clean
./autocti_assistant	main	clean
./autogalaxy_assistant	main	clean
./autofit_assistant	main	clean

Suggested branch: feature/science-project-memory
Worktree root: ~/Code/PyAutoLabs-wt/science-project-memory/

Implementation Steps

Phase 1 — PyAutoBrain/agents/conductors/clone/

1. _clone.py: add a sync subcommand path. Semantics — not a blind overwrite:
   • Resolve the reference (default autolens_assistant) and its _SHARED_GENERIC file set (reuse REFERENCE_PROFILES[...]["generic"] / match_any, not a second list).
   • Compute the reference's diff over a commit range (--since <rev>, default: the sibling's last recorded sync, else origin/main), restricted to the generic paths.
   • For each sibling, apply per file with git apply --3way --check then --3way, falling back to reporting rejected hunks. Report per file: applied / rejected (<hunk list>) / absent / unchanged.
   • Dry run by default (--check only, nothing written); --apply writes.
2. bin/pyauto-brain: wire clone sync ... through clone.sh.
3. agents/conductors/clone/AGENTS.md: document the mode, its semantics, and that a rejected hunk is a human decision, never auto-resolved.
4. tests/test_clone_sync.py: temp-dir fixture with three fake assistants (one divergent) asserting clean-apply, rejected-hunk reporting, absent-file reporting, and that a dry run writes nothing.

Phase 2 — autolens_assistant

• skills/start-new-project.md:
  • Generated project AGENTS.md template gains "## Session start — do this first, every session": read wiki/project/profile.md, wiki/project/state.md, and the newest dated wiki/project/YYYY-MM-DD-*.md, then answer.
  • Scaffold emits a project CLAUDE.md that @-imports AGENTS.md (the same two-line pattern as PyAutoLabs/CLAUDE.md).
  • Scaffold tree: add state.md + results_summary.md; remove results/manifests/ and every sentence promising per-run manifests (Phase 2 step 1, the generated AGENTS.md, the generated README.md, the Publish gate) — the journal run table is the run record.
  • Seed the project's profile.md from the assistant clone's own wiki/project/profile.md when present, else from _profile_template.md.
• wiki/project/_state_template.md (new, ~30 lines): Where we are now / In flight / Open, carried forward / Traps — don't repeat / Journal index.
• wiki/project/_results_summary_template.md (new): covers_through: <date> at the top.
• wiki/project/_profile_template.md: user-level sections only (background, interaction mode, HPC access + authorization, automation preference); science goal + data-on-hand move to the state template.
• wiki/project/README.md: state.md rule — rewritten each session, never appended; the end-of-session ritual becomes "journal entry + rewrite state.md"; an entry is finished when state.md is current.
• AGENTS.md step 2: read state.md and the newest journal entry too, when present.
• The clone boundary already covers wiki/project/* and skills/start-new-project* as _SHARED_GENERIC, so check_boundary.py stays green.

Phase 3 — propagation

• clone sync --reference autolens_assistant dry-run → all three siblings.
• Apply + PR autocti_assistant and autogalaxy_assistant if hunks land cleanly; in autocti_assistant reconcile the orphan wiki/project/state.md (real content: CTI stack/release-train/heritage facts) against the scaffolded template.
• autofit_assistant: no apply — post the dry-run report here.

Key Files

• PyAutoBrain/agents/conductors/clone/_clone.py — _SHARED_GENERIC, match_any, REFERENCE_PROFILES; new sync mode
• PyAutoBrain/agents/conductors/clone/AGENTS.md — operating summary
• PyAutoBrain/bin/pyauto-brain — conductor dispatch
• PyAutoBrain/tests/test_clone_sync.py — new
• autolens_assistant/skills/start-new-project.md — the scaffold (A, C, D)
• autolens_assistant/wiki/project/{README,_profile_template,_state_template,_results_summary_template}.md
• autolens_assistant/AGENTS.md — session-start step 2

Original Prompt

Click to expand starting prompt

Science-project memory: a fresh chat pointed at a project folder resumes…

Type: feature
Target: PyAutoBrain
Repos:

• autocti_assistant
• autofit_assistant
• autogalaxy_assistant
• autolens_assistant
• PyAutoBrain
  Difficulty: large
  Autonomy: supervised
  Priority: high
  Status: formalised

Science-project memory: a fresh chat pointed at a project folder resumes by default

Type: feature
Difficulty: large
Autonomy: supervised
Priority: high

Make "point a fresh agent chat at a science project folder and it works out where it got to, how to resume, and the project's history" true by default for projects scaffolded by the assistants' start-new-project skill.

Scope: autolens_assistant (reference copy), autofit_assistant, autogalaxy_assistant, autocti_assistant, and the pyautobrain clone conductor. The euclid assistant is a different design and is explicitly out of scope.

User's original request (verbatim): "I want to know how much memory of a project they truly have. Ideally, I can always load a new agent (e.g. claude) chat, and point it to a science project and it can quickly work out where it got to, how to resume and the general history of the science project. I feel like this would need more in the wiki than just the code run, so how much of this type of more persistant memory is avaialble? Is this something we should add, like an indexed tracker of tasks and work analogous to how organ repos and other tasks are easily resumeable?"

Findings verified 2026-08-28 (one real project scaffolded in July, two journal entries, one full run campaign):

1. The generated project AGENTS.md (from skills/start-new-project.md) has no session-start section; each assistant's own AGENTS.md step 2 reads wiki/project/profile.md only. Nothing reads the journal on arrival, so a fresh chat does not resume unless asked "what have we done?".
2. autocti_assistant/wiki/project/state.md is an orphan — no AGENTS.md or skill reads or writes it (the only precedent for a head pointer, and it is dead).
3. results_summary.md is consumed by the Publish phase (gh release create --notes-file wiki/project/results_summary.md, autolens_assistant skills/start-new-project.md:381) but is never templated or created by the scaffold.
4. results/manifests/ is aspirational: empty after a full real campaign; the journal's run tables do the job.
5. profile.md conflates user-level facts (role, HPC/SSH access, automation consent) with the per-project science goal; a second project re-elicits everything.
6. The four copies of skills/start-new-project.md and wiki/project/{README,_profile_template,_template}.md have four distinct hashes (the autofit_assistant copy diverged by 343 lines). The clone conductor lists them as _SHARED_GENERIC (agents/conductors/clone/_clone.py:69-81) but only has a birth mode — nothing re-syncs after birth.

Deliverables, in priority order (A alone gets most of the value):
A. Session-start block in the generated project AGENTS.md, and in the assistants' AGENTS.md step 2: read profile.md, wiki/project/state.md, and the newest dated journal entry before answering. Verify the scaffold emits a CLAUDE.md that @-imports AGENTS.md so Claude Code actually loads it.
B. wiki/project/state.md (keep autocti_assistant's filename) — a small head pointer, rewritten not appended each session: Where we are now / In flight (runs, job IDs, what unblocks each) / Open, carried forward (struck when done) / Traps — don't repeat / one-line journal index. Wire it into the existing end-of-session "want a journal entry?" ritual: an entry is finished when state.md is rewritten. Markdown read by an LLM, not YAML — a science project is one repo with a handful of in-flight runs.
C. Template results_summary.md with a covers_through: <date> stamp so staleness is visible. Drop results/manifests/ from the scaffold and the AGENTS text (delete the trap rather than document it).
D. Split profile.md at the user/project seam: the user half (who you are, HPC access, automation preference) seeded from the assistant clone's profile.md and free to diverge; the project goal moves to state.md.
E. Land via a clone-conductor sync mode that re-applies _SHARED_GENERIC files from autolens_assistant (reference) to the other three, rather than four hand edits. A human reviews the autofit_assistant divergence before any overwrite — some of it may be deliberate domain adaptation.

Out of scope: the library-repo anonymisation of a named science target is filed as its own small maintenance prompt in the data-reduction library.

[Jammy2211]

Branches pushed — PRs held at the Heart gate

All three phases are implemented and pushed. No PRs were opened: the ship gate consulted
pyauto-heart readiness --json and got RED. The reason strings, verbatim:

red_reasons:
  "release validation FAILED (stage integrate)"

yellow_reasons:
  "workspace validation not passing (2 failed, cloud#33179766004: autolens_test scripts/imaging/rectangular_mge.py, autolens_test scripts/imaging/rectangular_mge_rtu.py)"
  "manifest drift: session-start hooks (generated) — 32 mismatch(es) vs PyAutoMind/repos.yaml"

(verdict red, score 45, ts 2026-08-28T21:34:42Z, profile default, nothing stale.)

None of the RED reasons is caused by this work — it is markdown, one new conductor mode and a
new hermetic test file. Opening the four PRs is a one-command follow-up once Heart is GREEN,
or under a human-authorized corrective exception naming the RED reason above.

Repo	Branch	Commit	Tests
PyAutoBrain	feature/science-project-memory	55e9fd4	pytest tests/ -q 627 passed, 0 failed (618 pre-existing + 9 new in tests/test_clone_sync.py); tenant firewall OK
autolens_assistant	feature/science-project-memory	6b7c5c4	check_boundary complete (generic 60 · domain 324 · mixed 92)
autocti_assistant	feature/science-project-memory	523015b	check_boundary skips (not a reference)
autogalaxy_assistant	feature/science-project-memory	7964ff7	check_boundary skips (not a reference)

One supporting change is already on PyAutoMind main (2b764e48): the tenant-firewall
allowlist entry for the new Brain test. The Brain PR gate runs the firewall against Mind
main, so it had to land first.

Phase 1 — pyauto-brain clone sync

Not a blind overwrite. It takes the reference's own diff over a commit range, restricted to
that reference's generic pattern set (reusing REFERENCE_PROFILES/match_any), rewrites the
names for each sibling, and applies it with GNU patch. Per file it reports applied /
created / already-applied / rejected (hunks #n) / absent / skipped / unsupported.
Dry run by default, --apply writes and leaves .rej files; exit 1 when anything was
rejected. "Since the sibling's last sync" is read from the sibling's own history via a
Clone-sync: <reference>@<sha> commit trailer — no state file.

The substitution set adds the UPPERCASE rule birth omits, which is why autocti_assistant
still names $AUTOLENS_ASSISTANT in its generated project scaffold.

Phase 3 — autofit_assistant: dry run only, NOT applied

As planned, autofit_assistant was not written to. Its _SHARED_GENERIC copies have
diverged too far for a patch to be the right instrument, and some of that divergence is
deliberate domain adaptation (it is the domain-agnostic base, not a domain assistant). Its
dry-run report:

autofit_assistant  (since ee306ac — autolens_assistant origin/main)
  XX  AGENTS.md                                        rejected — hunks #1, #2 rejected
  OK  modes/assistant.md                               applied
  XX  skills/start-new-project.md                      rejected — hunks #2, #8, #9, #10 rejected
  XX  wiki/project/README.md                           rejected — hunks #1, #3, #4 rejected
  XX  wiki/project/_profile_template.md                rejected — hunks #1, #2 rejected
  NEW wiki/project/_results_summary_template.md        created
  NEW wiki/project/_state_template.md                  created
  summary: 1 applied · 2 created · 4 rejected

Divergence from the reference, after normalising the names (diff-changed lines):
AGENTS.md 193 · skills/start-new-project.md 221 · wiki/project/README.md 56 ·
wiki/project/_profile_template.md 101.

Left for a human. The two new template files would land cleanly; the four rejected files
need a decision per hunk about which side's prose wins. Re-run to reproduce:

bin/pyauto-brain clone sync --since ee306ac --target autofit_assistant

Also noted, deliberately not fixed here

• autocti_assistant/wiki/project/_profile_template.md still heads its first section
  "Lensing background", and its generated .claude/settings.json still names
  $AUTOLENS_ASSISTANT — birth-time substitution gaps that predate this work.
• autolens_assistant's llms-chat.txt and chat_pack/01_api_surface.md are already stale on
  main (chat_bundle.py --check fails there too). Regenerating them would sweep an unrelated
  API-surface snapshot into this PR, so it stays a separate hygiene item.

[Jammy2211]

Follow-up folded in: autofit_assistant synced, the two birth gaps closed at source

The follow-up prompt that was going to be its own issue is folded into this one — same
branches, no new issue. Human authorised both legs ("do this" on the autofit dry-run
report). Still no PRs: Heart is RED for the reasons already posted above, unchanged
and unrelated to this work.

Repo	Branch	Commit	Checks
PyAutoBrain	feature/science-project-memory	d2b1af9	pytest tests/ -q 634 passed (627 before; tests/test_clone_sync.py 9 → 16); tenant firewall OK
autocti_assistant	feature/science-project-memory	df8408d	check_boundary skips (not a reference); gap grep empty
autofit_assistant	feature/science-project-memory	1bc7675	pytest autoassistant/tests -q 56 passed; check_boundary complete (generic 105 · domain 20 · mixed 21)

autolens_assistant (6b7c5c4) and autogalaxy_assistant (7964ff7) are unchanged —
autogalaxy was checked for both gaps and carries neither (its profile template was
hand-adapted at birth). euclid_assistant was never opened.

Leg B — the gaps closed in the rename table, not by hand

Birth and sync each built their own substitution list, and birth's was missing the two
rules. name_substitutions() is now the one table both use — apply_seed serialises
it for PyAutoHands' clone_seed, sync_substitutions calls it directly — and it carries:

• the UPPERCASE package rule birth omitted ($AUTOLENS_ASSISTANT → $AUTOCTI_ASSISTANT);
• DOMAIN_NOUNS — the science's own noun and its longer phrase per package
  (lensing / galaxy / CTI / model-fitting; "gravitational lensing" → "charge transfer
  inefficiency"), word-anchored so microlensing survives and lensing-fluent keeps its tail;
• DOMAIN_ALIASES — the qualified compounds resolve to the target's bare noun first,
  because "strong-lensing" → "strong-CTI" is a phrase in no science.

A target package absent from the table gets no domain rule and a printed warning at
birth: a science's own noun is never guessed.

Seven new tests pin it, including the honest limit: sync does not retro-fix. It only
ever touches the lines in the patch, so a line an earlier birth got wrong stays wrong until
the reference changes that very line — which is why autocti needed the step below rather
than a re-run of clone sync (it was already synced to autolens_assistant@6b7c5c4, so a
re-run had nothing to replay).

autocti_assistant (df8408d) therefore took the conductor's own substitution output over
the reference's current copies of the two files — mechanical, not hand-written prose; the two
files were otherwise byte-identical to the substituted reference. Then three hand adaptations
in the profile template's examples, the way autogalaxy's were done at birth: a mechanical
rename cannot turn "PhD on weak lensing, new to strong" into a sentence a CTI user recognises.
Verified: grep -n 'AUTOLENS\|lensing\|Lensing\|autolens\|HowToLens' over both files is empty.

One extra find: .claude/skills/start-new-project.md in autocti was a stale copy of the
pre-#315 file (no session-start block, results/manifests/ still there) where every sibling
cell has a symlink — the same birth-artefact class. Restored as the symlink.

Leg A — autofit_assistant, the four rejected files

clone sync --since ee306ac --target autofit_assistant --apply reproduced the dry run
exactly (1 applied · 2 created · 4 rejected). Every reject resolved by the same rule:
autofit's domain adaptation wins, #315's structure is layered on. No .rej left.

• AGENTS.md (hunks 1, 2) — session-start step 2 becomes "Project memory" (state.md,
  newest entry, profile.md), keeping this cell's "and domain context"; the sub-wiki
  paragraph and the first-interaction protocol take the profile/state split; the
  domain-adaptation section now routes "where the work got to" to state.md.
• skills/start-new-project.md (hunks 2, 8, 9, 10) — results/manifests/ and the
  per-run manifest JSON give way to the journal run table + the state.md rewrite ritual,
  written in this cell's transit/TOI-1234 vocabulary rather than the reference's SLACS
  example; the HPC paragraph gains the user-vs-project split; Phase 3 starts an arriving
  collaborator from state.md. A stale "records the version in every run manifest" line
  further up was updated to the journal row.
• wiki/project/README.md (hunks 1, 3, 4) — the three-file memory, in this cell's own
  vocabulary (prior change, sampler swap, paper ingested).
• wiki/project/_profile_template.md (hunks 1, 2) — "User profile" + the portability
  note; "Science goal" and "Data inventory" dropped (they are state.md's first two
  sections). "Scientific domain" kept — a fact about the user, and this cell's most
  important section. "Likelihood & model code" moved to state.md — it is where this
  project's code sits, not who the user is.

modes/assistant.md applied clean. The two created templates were adapted for the
domain-agnostic base: _state_template.md gains the "Likelihood & model code" section, and
two lensing examples ("a mask radius that ate the arc", "pixelised source settings") became
domain-agnostic ones. Parity check against the reference: state.md mentioned 17 times in
each copy of the skill, results_summary 7 in each; no leftover results/manifests; the only
"autolens"/"lensing" strings left in autofit are the two genuine cross-references that were
already on main (it describes itself against a domain assistant).

Not fixed here — reported

• autocti's generic files still carry reference example strings no rename rule can reach:
  slacs_subhalo and a SLaM/slacs0946 run-table row in skills/start-new-project.md,
  slacs0737/subhalo filename examples in wiki/project/README.md, and lensing wording in
  its AGENTS.md wiki paragraph. These need the same human domain adaptation autogalaxy got;
  they are not substitutions.
• autocti has three more .claude/skills/ entries that are real files where the siblings
  have symlinks: _bootstrap_skill.md, _style.md, contribute-upstream.md. Same birth
  artefact, left alone here.
• autolens_assistant's llms-chat.txt / chat_pack/01_api_surface.md are still stale on
  main — unchanged from the note above.

[Jammy2211]

PRs open — under an acknowledged Heart RED

The human authorised opening and merging under the RED verbatim: "open prs under red and merge i acknowledge". The readiness reasons are unchanged from the comments above and none of them is caused by this work:

• RED: release validation FAILED (stage integrate)
• YELLOW: workspace validation not passing (2 failed, cloud#33179766004: autolens_test scripts/imaging/rectangular_mge.py, rectangular_mge_rtu.py)
• YELLOW: manifest drift: session-start hooks (generated) — 32 mismatch(es) vs PyAutoMind/repos.yaml

The acknowledgement is recorded on both task entries in PyAutoMind/active.md (heart-ack:).

Repo	PR	Branch head	Verification
PyAutoBrain	#316	d2b1af9	pytest tests/ -q 634 passed; tenant firewall OK
autolens_assistant	PyAutoLabs/autolens_assistant#116	6b7c5c4	check_boundary complete (generic 60 · domain 324 · mixed 92)
autogalaxy_assistant	PyAutoLabs/autogalaxy_assistant#20	7964ff7	check_boundary skips (not the reference); both birth gaps absent
autocti_assistant	PyAutoLabs/autocti_assistant#27	df8408d	check_boundary skips; gap grep empty
autofit_assistant	PyAutoLabs/autofit_assistant#31	1bc7675	pytest autoassistant/tests -q 56 passed; check_boundary complete (generic 105 · domain 20 · mixed 21)

The PyAutoBrain PR carries Closes #315; the four assistant PRs reference this issue as Part of. This issue closes once all five are merged.

Opened alongside, unrelated task, same acknowledgement: PyAutoLabs/PyAutoReduce#74 (PyAutoReduce#73).

[Jammy2211]

Three of five merged — reopening: two PRs are red on pre-existing wiki-currency drift

Repo	PR	CI	Result
PyAutoBrain	#316	Brain Tests — pytest (3.12) success, pytest (3.13) success	MERGED 8f5f7697658d585b42fa52ea6d6c0fcdd1988bda
autocti_assistant	#27	clone-boundary success, wiki-currency success	MERGED 0242d7be673673aa93c701d8b7ec37cbeba55566
autofit_assistant	#31	wiki-currency success	MERGED d5df11fba0bc2e9a54bf0318d705f0a4d0b3a1f1
autolens_assistant	#116	clone-boundary success · wiki-currency FAILURE	left open
autogalaxy_assistant	#20	clone-boundary success · wiki-currency FAILURE	left open

GitHub closed this issue automatically on #316's Closes #315. Reopened — two of its five PRs are still open.

The failing leg, verbatim

wiki-currency / job wiki-currency, step "Run set +e …":

##[error]wiki-currency drift detected — see the job summary / drift-report artifact.
##[error]Process completed with exit code 1.

autolens_assistant#116 (drift-report.md artifact, run 33226180760) — one sub-check failed:

## Symbol audit (--scope all)
[audit] skipped 2 marker-opted-out file(s): skills/al_audit_skill_apis.md, autoassistant/mcp/server.py
[audit] scanned 67 files; 141 unique symbols
[audit] missing/broken: 1
**FAILED** — `Symbol audit (--scope all)` exited 1

(--check-version clean, --lint-idioms clean, --check-provenance warnings only, --check-citations clean, chat bundle currency reported — the known stale llms-chat.txt / chat_pack/01_api_surface.md noted earlier in this thread.)

autogalaxy_assistant#20 (run 33226182888) — two sub-checks failed:

## Symbol audit (--scope all)
[audit] scanned 46 files; 220 unique symbols
[audit] missing/broken: 1
**FAILED** — `Symbol audit (--scope all)` exited 1

## Citation paths (--check-citations)
[citations] ERROR wiki/core/operations/sandbox.md: `PyAutoGalaxy:autogalaxy/plot/plot_utils.py` — autogalaxy/plot/plot_utils.py not in .../sources/PyAutoGalaxy
[citations] scanned 78 files, 1061 citation(s) (11 skipped) — 2 missing path(s), 1 warning(s).
**FAILED** — `Citation paths (--check-citations)` exited 1

Why this is not this branch's failure

Both failures are in wiki/core/ pages and the skill API audit. This branch touches only AGENTS.md, modes/assistant.md, skills/start-new-project.md, wiki/project/* and .claude/hooks/session-start.sh — no wiki/core page, no cited source path, no audited skill symbol.

wiki-currency was already failing on both repos' previous PRs, which were merged anyway:

• autolens_assistant: run 2026-08-23 on feature/pynufft-removal-residue-phase-2 → failure (PR refactor: rename morning skill to wake_up (composition door) #115, merged)
• autogalaxy_assistant: run 2026-08-23 on feature/pynufft-removal-residue-phase-2 → failure (PR Add /health full mode (release-run dashboard) #19, merged)

No code was modified to make the leg pass, and neither PR was merged. They stay open for a human: either re-run after the wiki drift is fixed as its own hygiene task, or merge deliberately on the same "pre-existing red" judgement the previous pair took.

[Jammy2211]

Shipped

All five PRs are merged. Phases 1-3 landed, plus the folded-in follow-up (autofit_assistant sync + the autocti_assistant birth-substitution gaps).

Repo	PR	Merge commit	What landed
PyAutoBrain	#316	8f5f769	Phase 1 — clone sync mode on the clone conductor (three-way patch of _SHARED_GENERIC from a reference cell), plus the shared name_substitutions() rename table used by both birth and sync (UPPERCASE package rule, DOMAIN_NOUNS / DOMAIN_ALIASES; unknown target science = no domain rule + a warning, never a guess). Suite 627 → 634; clone-sync tests 9 → 16.
autocti_assistant	#27	0242d7b	Phase 3 sync + the two birth-substitution gaps closed from the conductor's own output; orphan wiki/project/state.md reconciled into the template (not deleted); stale .claude/skills/start-new-project.md real-file copy restored as the symlink every sibling has.
autofit_assistant	#31	d5df11f	The folded-in leg — synced --since ee306ac, all four rejected files hand-resolved keeping its fitting/inference domain adaptation and layering #315's structure on top. 56 tests pass, boundary complete.
autolens_assistant	#116	010378d	Phase 2, the reference copy — deliverables A-D: session-start block in the generated project AGENTS.md (+ project CLAUDE.md that @-imports it), state.md head pointer (templated, scaffolded, rewritten-not-appended, wired into the end-of-session ritual), templated results_summary.md with covers_through:, results/manifests/ deleted from the scaffold and from every sentence promising it, profile.md split at the user/project seam.
autogalaxy_assistant	#20	5938bcc	Phase 3 sync — state.md as the head pointer a session reads first.

euclid_assistant was out of scope throughout and was never opened.

The last two merged over a pre-existing CI red

autolens_assistant#116 and autogalaxy_assistant#20 were merged on human acknowledgement with the wiki-currency leg red. The failure is pre-existing wiki/core / skill-audit drift that these branches do not touch:

• autolens_assistant#116 — "Symbol audit (--scope all) exited 1" (missing/broken: 1)
• autogalaxy_assistant#20 — the same, plus "Citation paths (--check-citations) exited 1" (wiki/core/operations/sandbox.md cites PyAutoGalaxy:autogalaxy/plot/plot_utils.py, absent from the source tree)

The same leg failed on both repos' previous PRs (#115, #19 — pynufft residue phase 2, 2026-08-23), which were merged anyway. No code was modified to make the leg pass.

That drift is now filed as its own task: PyAutoMind draft/maintenance/pyautobrain/fix_wiki_currency_ci_drift_in_the.md.

Heart was RED at ship time for an unrelated reason, acknowledged verbatim by the human: "release validation FAILED (stage integrate)".

Closing — the follow-up prompt was recorded as folded into this issue, not re-filed as a separate task.