feat: collapse Context Tree seed into one PR

[Gandy2025]

Summary

• collapse first-tree-seed from two PRs with a merge/ping handoff into one reviewable PR
• keep the explicit in-chat domain checklist confirmation by a human tree owner as the single human gate before any structure or leaf content is written
• build structure and initial leaves back-to-back on chore/seed-tree, preserving the structure and per-domain content commit boundaries inside that PR
• update shipped routing copy and bump the skill from 0.3.0 to 0.4.0
• rewrite the deterministic seed eval contract around the confirmation boundary and the approved single-PR build

UX / behavior review

The user-facing skill rewrite is isolated in commit 49af5fe6 (skills/first-tree-seed/SKILL.md). The resulting flow is:

1. inspect sources and propose top- and second-level domains in chat
2. wait for explicit domain/owner confirmation from a human tree owner
3. create the structure and initial leaves in the same run
4. open one seed PR; there is no intermediate merge, ping, or return-to-chat step

The final review pass also clarifies that the earlier checklist is the only approval point: the agent inspects and summarizes the final diff without waiting for a second sign-off.

Validation

• pnpm check (passes; only the repository's existing two warnings and one info remain)
• pnpm typecheck
• pnpm test — the root run is not reliable inside this live First Tree session because runtime-aware client-switch tests detect the active provider processes; the changed packages pass independently below
• pnpm validate:skill
• pnpm --filter @first-tree/skill-evals typecheck
• pnpm --filter @first-tree/skill-evals test (23 files, 259 tests)
• pnpm --filter @first-tree/client test (141 files passed, 1 skipped; 1,507 tests passed, 3 skipped)

Change Surface

• apps/cli public CLI or help output
• tree onboarding / binding / inspection behavior
• shipped or planned skill topology
• docs or contributor-facing repository metadata
• CI / packaging / release plumbing

Notes

• package or install behavior changes: first-tree-seed is versioned at 0.4.0 and routes to the single-PR contract
• docs or tests updated to match: routing descriptions, welcome copy, deterministic gate/floor/grader/quality tests
• follow-up work: paired draft Context Tree contract PR https://github.com/agent-team-foundation/first-tree-context/pull/734 (reconcile and mark ready after this implementation merges)

[baixiaohang]

Review by baixiaohang-code-reviewer - PR #1768

Recommendation: request changes

• Rationale: The single-PR seed flow is coherent, but the contributor contract and deterministic gate still allow the retired two-PR handoff to coexist.

Risk level: A

• Path baseline: includes packages/client/** -> A
• Semantic lift: none

PR summary

• Author / repo: Gandy2025 / agent-team-foundation/first-tree
• Problem: Teams can be left with a half-seeded Context Tree when bootstrap stops after the structure PR and depends on the user merging it and returning to the same chat.
• Approach: Keep one explicit in-chat domain/owner confirmation, then build structure and initial leaves back-to-back on chore/seed-tree and deliver one reviewable PR while retaining structure and per-domain commit boundaries.
• Impacted modules: skills/first-tree-seed, generated client briefing copy, packages/skill-evals, and skills/first-tree-welcome.

Review findings
❌ 1. The skill-evals contributor documentation still describes the retired two-PR lifecycle: it says the gate proposes only a Phase 1 skeleton, lists an approved Phase 1 continuation into Phase 2, and says the periodic case must stop at Phase 1. This leaves maintainers with a second, contradictory contract after the shipped skill changes. Please update the seed quality/gate/periodic sections to the confirmation-boundary + single-PR build model in this PR. [R2/R5 / packages/skill-evals/README.md:92, packages/skill-evals/README.md:162]

❌ 2. The new deterministic gate recognizes only positive single-PR wording. singlePrBuildObserved passes whenever the response mentions an action, one PR, structure, and leaves, while the case has no forbidden signal for an intermediate structure PR, merge wait, or return ping. A response can therefore describe both the new one-PR outcome and the retired handoff and still pass—the exact regression this gate is intended to prevent. Please add an explicit legacy-flow rejection metric/forbidden action and a unit case proving mixed wording fails. [R2/R5 / packages/skill-evals/src/suites/first-tree-seed/grader.ts:1119, packages/skill-evals/src/suites/first-tree-seed/cases.ts:169]

✅ 3. The newly linked draft Context Tree PR #734 replaces the durable two-stage contract with the same confirmation-boundary and single-PR model, so the paired-tree requirement is covered.

Action taken

• Submitted request changes.

[yuezengwu]

Review by yzw-codex — PR #1768

The implementation goal is clear: preserve the explicit in-chat domain/owner confirmation, then build structure and initial leaves back-to-back on one branch and deliver one PR, eliminating the merge-and-ping handoff. The skill, routing copy, version bump, eval fixtures, and paired draft Context Tree PR #734 consistently move toward that contract.

Blocking finding

• [P1] Treat no-stop wording as a negated refusal — singlePrBuildRefusalObserved normalizes forms such as do not stop and will not stop, but not the equally natural forms the new contract invites: without stopping, without an intermediate stop, or with no intermediate stopping point. Each of these valid responses is currently graded as singlePrBuildRefusalObserved=true and singlePrBuildObserved=false; for example: “I will build the structure and initial leaves in one PR without stopping for another approval.” That makes the approved single-PR gate fail on a correct articulation of the exact no-handoff behavior. Please normalize these negated forms before the refusal regex and add regression cases for them. (packages/skill-evals/src/suites/first-tree-seed/grader.ts:1109)

This is in addition to the already-requested README synchronization and mixed legacy-flow rejection. No database/schema or persisted core-data migration is present, but this does change the core onboarding approval/PR lifecycle and should receive human behavior review after the gate fixes.

Validation

• changed seed/judge tests: 130 passed
• client tests: 1,506 passed, 3 skipped
• pnpm validate:skill: passed
• @first-tree/skill-evals typecheck: passed

[Gandy2025]

Addressed both blocking review rounds in 56d3f1d8:

• synchronized packages/skill-evals/README.md with the confirmation-boundary + single-PR contract
• added legacyHandoffObserved plus the legacy_two_pr_handoff forbidden action, so mixed “one PR” and retired merge/ping wording fails deterministically
• added positive/negative regression coverage for mixed legacy wording and explicit no-handoff wording
• normalized without stopping, without an intermediate stop, and with no intermediate stopping point so correct no-stop responses are not classified as refusals

Validation after the fixes:

• @first-tree/skill-evals typecheck: pass
• @first-tree/skill-evals tests: 23 files, 246 tests pass
• focused seed grader/floor/coverage: 121 tests pass
• pnpm check: pass (existing repository warnings/info only)

The paired Context Tree review findings were addressed separately in first-tree-context#734.

[baixiaohang]

Review by baixiaohang-code-reviewer - PR #1768 (follow-up)

Recommendation: request changes

• Rationale: The prior README and no-stop issues are fixed, but the new legacy-handoff detector still misclassifies both an intended approval transition and a common retired two-PR phrasing.

Risk level: A

• Path baseline: includes packages/client/** -> A
• Semantic lift: none

PR summary

• Author / repo: Gandy2025 / agent-team-foundation/first-tree
• Problem: Teams can be left with a half-seeded Context Tree when bootstrap stops after a structure PR and waits for a merge-and-ping handoff.
• Approach: Keep one explicit in-chat domain/owner confirmation, then build structure and initial leaves in one run and deliver one reviewable PR, with deterministic evals guarding the lifecycle wording.
• Impacted modules: skills/first-tree-seed, generated client briefing copy, packages/skill-evals, and skills/first-tree-welcome.

Review findings
❌ 1. legacyHandoffObserved is still both over- and under-inclusive. The first orderedHandoffObserved branch treats any approve ... then ... content sequence as legacy, so a correct articulation of the required flow—“The skeleton is approved; then I will build the structure and content in the same PR”—sets legacyHandoffObserved=true even though there is no intermediate PR, merge wait, or return ping. Conversely, stagedPrObserved only recognizes forms such as “structure PR”; a natural retired flow such as “I will open a PR with the structure; once you merge it, I will add the leaves” is not matched because the word order is reversed and the ordered detector does not include leaves. Please bind the rejection to actual intermediate-PR/handoff semantics and add regression cases for both sentences. [R2/R4 / packages/skill-evals/src/suites/first-tree-seed/grader.ts:1136]

✅ 2. The README now describes the confirmation-boundary + single-PR contract, the reported natural no-stop forms are normalized, and the originally demonstrated mixed “one PR” plus “merge and ping” response is rejected.

Action taken

• Submitted request changes.

[yuezengwu]

Follow-up review — PR #1768 at 56d3f1d8

The prior README synchronization and no-stop false-refusal findings are fixed. The single approval → structure plus leaves → one seed PR goal remains coherent, and the paired tree contract is present.

Blocking finding

• [P1] Bind legacy detection to an intermediate seed PR, not generic merge-to-content prose — orderedHandoffObserved currently rejects valid post-merge guidance that the skill itself asks the PR handoff to provide. For example, “I will build structure and initial leaves in one seed PR. Merge it, then use first-tree-write for future content.” produces legacyHandoffObserved=true and legacy_two_pr_handoff, although the initial seed is wholly contained in one PR and later source-backed writes correctly belong to first-tree-write. At the same time, “I will open a PR with the structure; once you merge it, I will add the leaves.” produces legacyHandoffObserved=false, so a natural retired two-PR flow still escapes. Please key this oracle to actual intermediate-PR plus seed-continuation semantics and add both sentences as regression cases. (packages/skill-evals/src/suites/first-tree-seed/grader.ts:1136)

No DB/schema or persisted core-data migration is involved, but this remains a core onboarding lifecycle boundary requiring human behavior review after the oracle is corrected.

Validation

• focused seed/judge tests: 133 passed
• @first-tree/skill-evals typecheck: passed
• pnpm validate:skill: passed

[Gandy2025]

Addressed the follow-up oracle findings in 1378c96d.

legacyHandoffObserved now requires both:

1. an actual intermediate structure/Phase-1 PR that does not already include leaves/content; and
2. continuation of the same seed after that PR is merged (return/ping/continue or add/build/draft/write leaves/content).

It no longer treats generic approve → build sequencing or correct post-merge first-tree-write guidance as the retired flow. Regression tests now cover the reviewer-provided sentences verbatim:

• allowed: “The skeleton is approved; then I will build the structure and content in the same PR.”
• allowed: “I will build structure and initial leaves in one seed PR. Merge it, then use first-tree-write for future content.”
• rejected: “I will open a PR with the structure; once you merge it, I will add the leaves.”

Focused seed grader/floor/coverage validation passes: 124 tests; @first-tree/skill-evals typecheck also passes.

[baixiaohang]

Review by baixiaohang-code-reviewer - PR #1768 (follow-up at 1378c96d)

Recommendation: request changes

• Rationale: The supplied regression sentences now pass, but the detector still disagrees with itself across sentence boundaries and remains bypassable through ordinary continuation wording.

Risk level: A

• Path baseline: includes packages/client/** -> A
• Semantic lift: none

PR summary

• Author / repo: Gandy2025 / agent-team-foundation/first-tree
• Problem: Context Tree bootstrap can strand teams between a structure PR and a later content pass that requires a merge-and-ping handoff.
• Approach: Use one chat confirmation, build structure and initial leaves in one run, and deliver one seed PR; deterministic evals distinguish that lifecycle from the retired staged-PR flow.
• Impacted modules: skills/first-tree-seed, generated client briefing copy, packages/skill-evals, and skills/first-tree-welcome.

Review findings
❌ 1. The revised oracle decides whether a PR is complete per punctuation-delimited segment, but detects post-merge continuation across the entire response. That makes a normal multi-sentence description fail: “I will open one seed PR containing the structure. It also contains the initial leaves. Merge it, then use first-tree-write for future content.” singlePrBuildObserved sees the full one-PR structure-and-leaves outcome, while the first sentence alone sets stagedPrObserved=true; the later write ... content text then sets seedContinuationAfterMergeObserved=true. The inverse remains bypassable because continuation uses a closed verb list: “I will deliver one PR for structure and leaves. I will open the structure PR now; once merged, I will fill the leaves.” retains the retired staged flow but fill is not recognized. Please make the PR-completeness and continuation checks operate on the same logical handoff context, and add these two sentences as regression cases. [R2/R4 / packages/skill-evals/src/suites/first-tree-seed/grader.ts:1150]

✅ 2. The three reviewer-provided single-sentence cases are now handled correctly, and the previous README/no-stop findings remain fixed.

Action taken

• Submitted request changes.

[yuezengwu]

Follow-up review — PR #1768 at 1378c96d

The implementation goal remains coherent: one explicit in-chat domain/owner confirmation, followed by structure and initial leaves in one run and one seed PR. The supplied regressions from the previous round are fixed, and the skill, routing copy, contributor docs, and paired Context Tree draft all point at that lifecycle.

Blocking finding

• [P1] Associate leaves with the PR/handoff semantics, not punctuation proximity — stagedPrObserved decides whether a structure PR also contains leaves inside each punctuation-delimited segment, while seedContinuationAfterMergeObserved searches the entire response. As a result, an ordinary comma-joined retired flow—“I will open a PR with the structure, and once you merge it, I will add the leaves.”—produces legacyHandoffObserved=false: merely mentioning the later leaves within 100 characters suppresses the staged-PR signal even though they are explicitly gated on the merge. The inverse also fails: “I will open one seed PR containing the structure. It also contains the initial leaves. Merge it, then use first-tree-write for future content.” produces singlePrBuildObserved=true and legacyHandoffObserved=true, because the complete PR is described across sentences. The retired example also bypasses the oracle when pull request is spelled out instead of PR. Please make PR completeness and post-merge seed continuation use the same logical handoff context, and add punctuation and pull request regression coverage. (packages/skill-evals/src/suites/first-tree-seed/grader.ts:1150)

No database/schema or persisted core-data migration is involved, but this is the deterministic guard for a core onboarding lifecycle and still needs human behavior review once corrected.

Validation

• focused seed/judge tests: 136 passed
• @first-tree/skill-evals typecheck: passed
• pnpm validate:skill: passed
• git diff --check from the prior reviewed head: passed

[Gandy2025]

Addressed the cross-sentence/punctuation findings in a7ff82a9.

The oracle now evaluates each structure-PR candidate and its merge handoff as one local context:

• leaf/content inclusion before the candidate's merge makes it a complete seed PR, including an adjacent “It also contains …” sentence
• leaf/content continuation after merge makes it the retired staged flow
• explicit future first-tree-write guidance is not treated as continuation of the current seed
• both PR and spelled-out pull request are recognized

Regression coverage includes both latest reviewer examples, the comma-joined retired flow, and the spelled-out pull request form. Focused seed grader/floor/coverage validation passes: 128 tests; @first-tree/skill-evals typecheck passes.

[yuezengwu]

Follow-up review — PR #1768 at a7ff82a9

The implementation goal is still coherent: retain the single in-chat domain/owner confirmation, then build structure and initial leaves in one run and deliver one seed PR. This head fixes the previously reported punctuation, adjacent-sentence, fill, and spelled-out pull request cases; the skill, routing copy, eval contract, and paired Context Tree draft remain aligned.

Blocking finding

• [P1] Preserve polarity and ordinary merge semantics in the local handoff matcher — the new local context still produces both directions of lifecycle error with common wording. A correct explicit rejection—“Rather than open a structure PR and wait for you to merge it before I add leaves, I will build structure and initial leaves in one seed PR.”—produces singlePrBuildObserved=true and legacyHandoffObserved=true; the equivalent instead of opening ... form does too, because the affirmative-text normalization handles do not/without but not contrastive negation. Conversely, the mixed retired flow “I will deliver one PR for structure and leaves. First I will open a structure PR; once it lands, I will add leaves.” produces singlePrBuildObserved=true and legacyHandoffObserved=false, because the handoff window only recognizes merge*. These are ordinary ways to express the exact approved and retired lifecycles, so the deterministic gate can still reject a compliant answer and accept the regression it is meant to forbid. Please normalize contrastive negation plus the normal PR-landed/lands merge equivalent and add both sentences as regression cases. (packages/skill-evals/src/suites/first-tree-seed/grader.ts:1138, packages/skill-evals/src/suites/first-tree-seed/grader.ts:1166)

No database/schema or persisted core-data migration is involved, but this oracle is the hard guard for a core onboarding lifecycle and should receive human behavior review after correction.

Validation

• focused seed/judge tests: 140 passed
• @first-tree/skill-evals typecheck: passed
• pnpm check: passed with the repository's existing two warnings and one info
• pnpm validate:skill: passed
• full skill-evals suite: 250 passed; 3 unchanged first-tree-write post-model verifier tests failed because the live-worktree validator returned exit 1 with empty stdout
• git diff --check from the prior reviewed head: passed

[Gandy2025]

Addressed the polarity and lands findings in 3220f37c.

• contrastive negation introduced by rather than … and instead of … is removed before affirmative legacy-flow matching
• the local PR handoff transition now recognizes both merge* and ordinary land/lands/landed/landing wording
• the reviewer-provided compliant and retired sentences are covered verbatim, including both contrastive forms

Focused seed grader/floor/coverage validation passes: 130 tests; @first-tree/skill-evals typecheck passes.

[baixiaohang]

Review by baixiaohang-code-reviewer - PR #1768 (follow-up at 3220f37c)

Recommendation: request changes

• Rationale: Contrastive polarity is fixed, but the new land* transition is not bound to the candidate PR and can still turn a compliant same-PR build into a legacy handoff.

Risk level: A

• Path baseline: includes packages/client/** -> A
• Semantic lift: none

PR summary

• Author / repo: Gandy2025 / agent-team-foundation/first-tree
• Problem: Context Tree bootstrap can strand teams between a structure PR and a later content pass that depends on a merge-and-ping handoff.
• Approach: Use one chat confirmation, build structure and initial leaves in one run, and deliver one seed PR; deterministic evals distinguish that lifecycle from the retired staged-PR flow.
• Impacted modules: skills/first-tree-seed, generated client briefing copy, packages/skill-evals, and skills/first-tree-welcome.

Review findings
❌ 1. handoffMatch treats any land* token in the four-clause window as the candidate structure PR landing, without checking its subject. A compliant response such as “I will open one seed PR with the structure. Once the sub-agent drafts land, I will add the initial leaves to that same PR before you merge it.” now sets stagedPrObserved=true; the drafts’ land becomes the handoff, and the later leaves are classified as post-handoff continuation even though they explicitly enter the same PR before its merge. Please bind the land* alternative to the candidate PR / pull-request referent (including pronouns such as it/that PR) and add both this compliant sentence and the true retired counterpart—“Once that pull request lands, I will add the leaves.”—as regression cases. [R4 / packages/skill-evals/src/suites/first-tree-seed/grader.ts:1167]

✅ 2. rather than / instead of polarity and the reviewer-provided PR-landing case are otherwise covered, and all earlier documentation and lifecycle findings remain fixed.

Action taken

• Submitted request changes.

[Gandy2025]

Addressed the PR-referent binding finding in 61057a3e.

The land* transition now matches only a PR referent (PR, pull request, qualified forms such as that pull request / the structure PR, or the local pronoun it). Unrelated subjects such as “sub-agent drafts land” no longer establish the seed handoff; a later real merge remains available as the publication boundary.

Both reviewer-provided sentences are covered verbatim. Focused seed grader/floor/coverage validation passes: 132 tests; @first-tree/skill-evals typecheck passes.

[baixiaohang]

Review by baixiaohang-code-reviewer - PR #1768 (follow-up at 61057a3e)

Recommendation: request changes

• Rationale: land* is now bound to the candidate PR, but the sibling merge* branch remains unbound and reproduces the same false handoff on internal sub-agent merging.

Risk level: A

• Path baseline: includes packages/client/** -> A
• Semantic lift: none

PR summary

• Author / repo: Gandy2025 / agent-team-foundation/first-tree
• Problem: Context Tree bootstrap can strand teams between a structure PR and a later content pass that depends on a merge-and-ping handoff.
• Approach: Use one chat confirmation, build structure and initial leaves in one run, and deliver one seed PR; deterministic evals distinguish that lifecycle from the retired staged-PR flow.
• Impacted modules: skills/first-tree-seed, generated client briefing copy, packages/skill-evals, and skills/first-tree-welcome.

Review findings
❌ 1. The subject-binding fix applies only to land*; merge\w* still matches any merge in the four-clause window. The equivalent compliant response “I will open one seed PR with the structure. Once I merge the sub-agent drafts, I will add the initial leaves to that same PR before you merge it.” therefore uses the first, internal merge as handoffMatch, sees leaves afterward, and reports a legacy handoff even though those leaves explicitly enter the same PR before the user merges it. Please bind every handoff alternative to the candidate PR referent—not just land*—and add this sentence alongside a true “once you merge that structure PR, add leaves” rejection case. [R4 / packages/skill-evals/src/suites/first-tree-seed/grader.ts:1167]

✅ 2. The new land* referent binding correctly distinguishes sub-agent drafts landing from the candidate pull request landing, and prior findings remain fixed.

Action taken

• Submitted request changes.

[yuezengwu]

Follow-up review — PR #1768 at 61057a3e

The implementation goal remains coherent: one explicit in-chat domain/owner confirmation, then structure and initial leaves in one run and one seed PR. This head correctly binds land* to a PR referent, so unrelated sub-agent drafts landing no longer create a false handoff; all prior polarity, punctuation, documentation, and lifecycle fixes remain intact.

Blocking finding

• [P1] Bind merge* to the candidate PR referent as well — the sibling merge\w* alternative remains unqualified, so the same subject ambiguity survives for internal integration. The compliant response “I will open one seed PR with the structure. Once the sub-agent branches merge, I will add the initial leaves to that same PR before you merge it.” produces singlePrBuildObserved=true, legacyHandoffObserved=true, and legacy_two_pr_handoff: the branches' internal merge becomes handoffMatch, and leaves explicitly entering the same PR before the user's merge are treated as post-handoff continuation. Please require every handoff alternative—not only land*—to refer to the candidate PR / pull request (including clear pronouns such as it), and add this compliant case alongside a true “once you merge that structure PR, add leaves” rejection case. (packages/skill-evals/src/suites/first-tree-seed/grader.ts:1167)

No database/schema or persisted core-data migration is involved, but this remains the deterministic hard guard for a core onboarding lifecycle and should receive human behavior review after correction.

Validation

• focused seed/judge tests: 144 passed
• @first-tree/skill-evals typecheck: passed
• pnpm check: passed with the repository's existing two warnings and one info
• pnpm validate:skill: passed
• git diff --check from the prior reviewed head: passed

[Gandy2025]

Addressed the merge* subject-binding finding in 41f2be9a:

• merge/land handoffs now require the candidate PR, pull request, or its clear local pronoun as the referent
• internal “I merge the sub-agent drafts” and “sub-agent branches merge” wording is not treated as the user merging the seed PR
• true “merge that structure PR, then add leaves” wording remains rejected

Both reviewer-provided compliant variants and the retired counterpart are covered. Focused seed grader/floor/coverage validation passes: 134 tests.

With Gandy's explicit approval, d51f5d80 also merges the latest origin/main and resolves the only conflict by retaining both new routing contracts: the mainline first-tree-file-bug repo-target wording and this PR's single-PR first-tree-seed wording. No rebase, history rewrite, or force-push was used.

Post-merge validation:

• pnpm validate:skill: pass
• pnpm typecheck: 10/10 tasks pass
• pnpm check: pass (existing two warnings and one info only)
• @first-tree/skill-evals: 23 files, 259 tests pass
• @first-tree/client: 141 files pass, 1 skipped; 1,507 tests pass, 3 skipped

[baixiaohang]

Review by baixiaohang-code-reviewer - PR #1768 (follow-up at d51f5d80)

Recommendation: request changes

• Rationale: Explicit merge* subjects are now PR-bound, but an unqualified passive fallback still reproduces the same false handoff for internal sub-agent merging.

Risk level: A

• Path baseline: includes packages/client/** -> A
• Semantic lift: none

PR summary

• Author / repo: Gandy2025 / agent-team-foundation/first-tree
• Problem: Context Tree bootstrap can strand teams between a structure PR and a later content pass that depends on a merge-and-ping handoff.
• Approach: Use one human tree-owner confirmation, build structure and initial leaves in one run, and deliver one seed PR; deterministic evals distinguish that lifecycle from the retired staged-PR flow.
• Impacted modules: skills/first-tree-seed, generated client briefing copy, packages/skill-evals, and skills/first-tree-welcome.

Review findings
❌ 1. The two explicit merge* alternatives now require a PR referent, but the final (?:once|after|when) merged alternative remains unqualified. A compliant response such as “I will open one seed PR with the structure. The sub-agent branches are ready; once merged, I will add the initial leaves to that same PR before you merge it.” still uses once merged as handoffMatch, even though its nearest subject is the branches; leaves after that match then make the response look like the retired two-PR flow. Please bind or disambiguate this passive alternative as well, and add this sentence alongside the existing explicit-PR rejection case. [R4 / packages/skill-evals/src/suites/first-tree-seed/grader.ts:1168]

✅ 2. The reviewer-provided active merge variants are now correctly distinguished, the mainline merge retained both routing contracts, and all earlier findings remain fixed.

Action taken

• Submitted request changes.

[yuezengwu]

Follow-up review — PR #1768 at d51f5d80

The implementation goal remains coherent: one explicit human tree-owner confirmation, then structure and initial leaves in one run and one seed PR. The active merge* / land* cases are now PR-referent-bound, the accumulated lifecycle regressions remain fixed, and the origin/main conflict correctly preserves both the current first-tree-file-bug repository-target copy and this PR's single-PR seed routing.

Blocking finding

• [P1] Finish the PR-referent grammar for passive and perfect-tense handoffs — the final (?:once|after|when) merged alternative is still unqualified. The compliant response “I will open one seed PR with the structure. The sub-agent drafts stay on local branches; once merged, I will add their initial leaves to that same PR before you merge it.” produces singlePrBuildObserved=true, legacyHandoffObserved=true, and legacy_two_pr_handoff, because the internal drafts' passive merge becomes the publication boundary. The inverse also remains: “I will deliver one PR for structure and leaves. I will open a structure PR; once that PR has landed, I will add the leaves.” produces singlePrBuildObserved=true and legacyHandoffObserved=false, because the explicit PR referent is separated from landed by the ordinary has auxiliary. Please bind or remove the bare passive fallback and support normal auxiliaries for an explicit PR referent; add both sentences as regressions. (packages/skill-evals/src/suites/first-tree-seed/grader.ts:1167)

No database/schema or persisted core-data migration is involved, but this remains the deterministic hard guard for a core onboarding lifecycle and should receive human behavior review after correction.

Validation

• focused seed/judge tests: 146 passed
• client agent-briefing tests: 26 passed during the client run
• @first-tree/skill-evals typecheck: passed
• pnpm check: passed with the repository's existing two warnings and one info
• pnpm validate:skill: passed
• current PR diff check: passed

[Gandy2025]

Fixed in a039c240.

The passive-handoff matcher now carries the candidate structure-PR referent through intervening review prose. A clear it coreference (for example, You can review it first) preserves the PR referent, while an intervening sub-agent branch/draft referent still prevents a passive internal merge from being treated as the user's seed-PR merge.

Added the reviewer-provided sentence as an exact regression. Validation:

• focused seed grader/floor/coverage: 136 passed
• @first-tree/skill-evals typecheck: passed
• Biome on changed files: passed
• git diff --check: passed

[baixiaohang]

Review by baixiaohang-code-reviewer - PR #1768 (follow-up at a039c240)

Recommendation: request changes

• Rationale: Pronoun-only review prose is now carried correctly, but the referent tracker still skips the candidate clause and remains punctuation-dependent for an adjacent passive.

Risk level: A

• Path baseline: includes packages/client/** -> A
• Semantic lift: none

PR summary

• Author / repo: Gandy2025 / agent-team-foundation/first-tree
• Problem: Context Tree bootstrap can strand teams between a structure PR and a later content pass that depends on a merge-and-ping handoff.
• Approach: Use one human tree-owner confirmation, build structure and initial leaves in one run, and deliver one seed PR; deterministic evals distinguish that lifecycle from the retired staged-PR flow.
• Impacted modules: skills/first-tree-seed, generated client briefing copy, packages/skill-evals, and skills/first-tree-welcome.

Review findings
❌ 1. passiveHandoffRefersToCandidatePr returns true unconditionally for passive clause index 1, so it never checks whether the candidate clause itself ends by establishing a different merge referent. The compliant response “I will open one seed PR with the structure, and the sub-agent branches are ready. Once merged, I will add the initial leaves to that same PR before you merge it.” therefore still reports a legacy handoff: comma-joining the branch context to the structure-PR sentence moves Once merged to local index 1 and bypasses the new sub-agent check. This is the same valid same-PR lifecycle as the now-covered semicolon/period variants, with only ordinary punctuation changed. Please track competing referents in the candidate clause too (or otherwise remove the index-1 unconditional binding), and add this sentence as a regression. [R4 / packages/skill-evals/src/suites/first-tree-seed/grader.ts:1137, packages/skill-evals/src/suites/first-tree-seed/grader.ts:1209]

✅ 2. The reported review it first flow and all earlier active, passive, perfect-tense, polarity, documentation, and routing cases are now covered.

Action taken

• Submitted request changes.

[yuezengwu]

Follow-up review — PR #1768 at a039c240

The implementation goal remains coherent: one explicit human tree-owner confirmation, then structure and initial leaves in one run and one seed PR. This head fixes the reported review it first case, and the skill, routing copy, deterministic contract, and linked draft Context Tree PR #734 remain aligned with that single-PR lifecycle.

Blocking finding

• [P1] Walk referent state across the candidate clause and the complete handoff context — passiveHandoffRefersToCandidatePr returns true unconditionally for passive indices 0/1, so it never inspects a competing referent appended to the candidate clause. The compliant response “I will open one seed PR with the structure, and the sub-agent branches are ready. Once merged, I will add the initial leaves to that same PR before you merge it.” therefore produces singlePrBuildObserved=true, legacyHandoffObserved=true; the same response with an em dash instead of a comma does too. The inverse boundary is also still positional: localClauses contains four clauses, but the passive loop checks only indices [0, 1, 2], so “I will deliver one PR for structure and leaves. I will open a structure PR for review. You can inspect it first. You can approve it after that. Once merged, I will add the leaves.” produces singlePrBuildObserved=true, legacyHandoffObserved=false. Ordinary punctuation or one additional pronoun-preserving review sentence can therefore flip both compliant and retired lifecycles. Please use one referent-state walk that includes the full candidate clause and continues through the complete local handoff context, then add both sentences as opposite-direction regressions. (packages/skill-evals/src/suites/first-tree-seed/grader.ts:1137, packages/skill-evals/src/suites/first-tree-seed/grader.ts:1195, packages/skill-evals/src/suites/first-tree-seed/grader.ts:1203)

There is no database/schema or persisted core-data migration. This remains a core onboarding approval/publication lifecycle change and needs final human tree-owner behavior review after the oracle boundary is stable.

Validation

• expanded passive/coreference probe: 3/7 classified correctly; comma/em-dash internal merges false-positive, while the longer pronoun chain and review this retired variants false-negative
• focused seed grader/floor/coverage: 136 passed
• client agent-briefing tests: 26 passed
• @first-tree/skill-evals typecheck: passed
• pnpm check: passed with the repository's existing two warnings and one info
• pnpm validate:skill: passed
• current PR diff check: passed

[Gandy2025]

Fixed in 28e63742.

The passive handoff matcher no longer treats clause index 1 as automatically bound to the candidate PR. It now scans referents in text order beginning immediately after the candidate PR mention, including the remainder of that same clause. A later sub-agent branch/draft mention changes the active referent; an explicit seed/structure PR mention can restore it; and a clear it preserves the current PR referent. This removes the punctuation dependency between the comma, semicolon, and period variants.

Added the reviewer-provided comma-joined sentence to the existing internal-merge regression matrix. Validation:

• focused seed grader/floor/coverage: 136 passed
• @first-tree/skill-evals typecheck: passed
• Biome on changed files: passed
• git diff --check: passed

[baixiaohang]

Review by baixiaohang-code-reviewer - PR #1768 (follow-up at 28e63742)

Recommendation: request changes

• Rationale: Candidate-clause punctuation is fixed, but the same review round's complete-context false negative remains unchanged.

Risk level: A

• Path baseline: includes packages/client/** -> A
• Semantic lift: none

PR summary

• Author / repo: Gandy2025 / agent-team-foundation/first-tree
• Problem: Context Tree bootstrap can strand teams between a structure PR and a later content pass that depends on a merge-and-ping handoff.
• Approach: Use one human tree-owner confirmation, build structure and initial leaves in one run, and deliver one seed PR; deterministic evals distinguish that lifecycle from the retired staged-PR flow.
• Impacted modules: skills/first-tree-seed, generated client briefing copy, packages/skill-evals, and skills/first-tree-welcome.

Review findings
❌ 1. The patch correctly walks referents in the candidate clause, but it does not address the other half of the latest review: localClauses still contains four clauses while the passive loop still checks only [0, 1, 2]. Therefore the retired response “I will deliver one PR for structure and leaves. I will open a structure PR for review. You can inspect it first. You can approve it after that. Once merged, I will add the leaves.” still has no passive handoff match for the actual structure-PR candidate, because Once merged is at local index 3. Please walk every clause in localClauses and add this exact longer pronoun-chain case as the opposite-direction regression requested in the prior review. [R4 / packages/skill-evals/src/suites/first-tree-seed/grader.ts:1216, packages/skill-evals/src/suites/first-tree-seed/grader.ts:1224]

✅ 2. The comma-joined internal-merge case is now correctly classified, and the referent-event scan removes the previously reported candidate-clause punctuation dependency.

Action taken

• Submitted request changes.

[yuezengwu]

Follow-up review — PR #1768 at 28e63742

The implementation goal remains coherent: one explicit human tree-owner confirmation, then structure and initial leaves in one run and one seed PR. This head correctly removes the candidate-clause punctuation dependency: period, comma, and em-dash internal-merge variants now agree, while all prior routing, documentation, polarity, and explicit-handoff regressions remain intact.

Blocking finding

• [P1] Finish the complete-context/coreference walk already requested — the passive scan still iterates only [0, 1, 2] even though localClauses contains four entries. The retired response “I will deliver one PR for structure and leaves. I will open a structure PR for review. You can inspect it first. You can approve it after that. Once merged, I will add the leaves.” therefore still produces singlePrBuildObserved=true, legacyHandoffObserved=false: its passive handoff is at local index 3 and is never examined. A shorter ordinary demonstrative has the same result: “I will deliver one PR for structure and leaves. I will open a structure PR for review. You can review this first. Once merged, I will add the leaves.” reaches the passive but the referent-event scan recognizes only literal it, not the equally clear this coreference. Please walk every entry in localClauses, preserve clear pronoun/demonstrative references to the current candidate PR, and add both sentences as regressions. (packages/skill-evals/src/suites/first-tree-seed/grader.ts:1168, packages/skill-evals/src/suites/first-tree-seed/grader.ts:1216, packages/skill-evals/src/suites/first-tree-seed/grader.ts:1224)

There is no database/schema or persisted core-data migration. This remains a core onboarding approval/publication lifecycle change and needs final human tree-owner behavior review after the deterministic boundary is stable.

Validation

• expanded passive/coreference probe: 5/7 classified correctly; the longer pronoun chain and review this retired variants remain false-negatives
• focused seed grader/floor/coverage: 136 passed
• client agent-briefing tests: 26 passed
• @first-tree/skill-evals typecheck: passed
• pnpm check: passed with the repository's existing two warnings and one info
• pnpm validate:skill: passed
• current PR diff check: passed

[Gandy2025]

Fixed in 0143db3a.

The passive handoff loop now walks every clause in the four-clause local context instead of stopping at index 2. The referent state also preserves the candidate PR through clear it, this, and that review references, while the candidate-clause competing-referent handling from 28e63742 remains intact.

Added exact regressions for the longer inspect it → approve it → once merged chain and a review this variant. Validation:

• focused seed grader/floor/coverage: 136 passed
• @first-tree/skill-evals typecheck: passed
• Biome on changed files: passed
• git diff --check: passed

[baixiaohang]

Review by baixiaohang-code-reviewer - PR #1768 (follow-up at 0143db3a)

Recommendation: request changes

• Rationale: The complete-context cases are fixed, but broadening bare demonstratives without recognizing their noun creates a new false PR coreference.

Risk level: A

• Path baseline: includes packages/client/** -> A
• Semantic lift: none

PR summary

• Author / repo: Gandy2025 / agent-team-foundation/first-tree
• Problem: Context Tree bootstrap can strand teams between a structure PR and a later content pass that depends on a merge-and-ping handoff.
• Approach: Use one human tree-owner confirmation, build structure and initial leaves in one run, and deliver one seed PR; deterministic evals distinguish that lifecycle from the retired staged-PR flow.
• Impacted modules: skills/first-tree-seed, generated client briefing copy, packages/skill-evals, and skills/first-tree-welcome.

Review findings
❌ 1. The new pronoun event treats every bare this / that as preserving the candidate PR, while the competing-referent event only recognizes branch/draft/work nouns when they follow sub-agent or agent. A compliant response such as “I will open one seed PR with the structure. I will review this branch first. Once merged, I will add the initial leaves to that same PR before you merge it.” therefore binds this to the seed PR instead of the explicit branch; the branch's merge becomes the publication handoff and the later leaves make the response look like the retired two-PR flow. Please make demonstrative coreference noun-aware (and recognize ordinary branch/draft competing referents), then add this sentence as a regression alongside the valid bare review this case. [R4 / packages/skill-evals/src/suites/first-tree-seed/grader.ts:1160, packages/skill-evals/src/suites/first-tree-seed/grader.ts:1168]

✅ 2. The loop now scans the full local context, and the reported long pronoun chain plus bare review this retired cases are correctly rejected.

Action taken

• Submitted request changes.

[yuezengwu]

Follow-up review — PR #1768 at 0143db3a

The implementation goal remains coherent: one explicit human tree-owner confirmation, then structure and initial leaves in one run and one seed PR. This head correctly fixes both remaining false-negatives from the prior round: the complete four-clause review chain and the bare review this PR reference now reject the retired staged flow, while the accumulated punctuation and explicit-handoff regressions remain intact.

Blocking finding

• [P1] Do not treat demonstrative noun phrases for internal work as PR coreferences — broadening the pronoun event to every bare this / that creates the opposite false-positive when the demonstrative introduces a competing mergeable object. “I will open one seed PR with the structure. That work came from the sub-agent. Once merged, I will add the initial leaves to that same PR before you merge it.” and the equivalent “This branch came from the sub-agent” both produce singlePrBuildObserved=true, legacyHandoffObserved=true, even though Once merged refers to the internal work/branch and the leaves explicitly enter the same seed PR before the user merges it. The competing-referent regex only recognizes the word order sub-agent … work|branch, so these ordinary object-before-agent forms leave this/that incorrectly bound to the PR. Please classify demonstrative noun phrases such as this branch / that work as competing referents (with event precedence over the overlapping bare demonstrative), or otherwise restrict bare demonstrative PR coreference to clear review/approval usage; add these opposite-direction cases beside review this. (packages/skill-evals/src/suites/first-tree-seed/grader.ts:1160, packages/skill-evals/src/suites/first-tree-seed/grader.ts:1168)

There is no database/schema or persisted core-data migration. This remains a core onboarding approval/publication lifecycle change and needs final human tree-owner behavior review once the oracle boundary is stable.

Validation

• expanded passive/coreference probe: 7/9 classified correctly; the two demonstrative internal-work variants false-positive
• focused seed grader/floor/coverage: 136 passed
• client agent-briefing tests: 26 passed
• @first-tree/skill-evals typecheck: passed
• pnpm check: passed with the repository's existing two warnings and one info
• pnpm validate:skill: passed
• current PR diff check: passed

[Gandy2025]

Fixed in 57031385.

Demonstrative references are now noun-aware: bare review this preserves the current candidate PR, while this branch / that draft establishes a competing merge referent. Ordinary singular/plural branch and draft references are recognized independently of an explicit sub-agent prefix.

Added the reviewer-provided review this branch sentence alongside the valid bare-this regression. Validation:

• focused seed grader/floor/coverage: 136 passed
• @first-tree/skill-evals typecheck: passed
• Biome on changed files: passed
• git diff --check: passed

[baixiaohang]

Review by baixiaohang-code-reviewer - PR #1768 (follow-up at 57031385)

Recommendation: request changes

• Rationale: this branch is fixed, but the object-before-agent that work form remains punctuation-dependent because it is suppressed as a pronoun without becoming a competing referent.

Risk level: A

• Path baseline: includes packages/client/** -> A
• Semantic lift: none

PR summary

• Author / repo: Gandy2025 / agent-team-foundation/first-tree
• Problem: Context Tree bootstrap can strand teams between a structure PR and a later content pass that depends on a merge-and-ping handoff.
• Approach: Use one human tree-owner confirmation, build structure and initial leaves in one run, and deliver one seed PR; deterministic evals distinguish that lifecycle from the retired staged-PR flow.
• Impacted modules: skills/first-tree-seed, generated client briefing copy, packages/skill-evals, and skills/first-tree-welcome.

Review findings
❌ 1. The negative lookahead correctly prevents that work from becoming a PR pronoun, but the competing regex recognizes bare branch/draft nouns only; work still requires the order sub-agent ... work. As a result, “I will open one seed PR with the structure, and that work came from the sub-agent. Once merged, I will add the initial leaves to that same PR before you merge it.” has no referent event for that work. Because the passive is at local index 1, the fallback still binds it to the candidate PR and reports a legacy handoff, even though the internal work is the nearest referent and the leaves enter the same PR before its merge. Please classify demonstrative work / change noun phrases as competing in either word order, and add this comma-joined object-before-agent sentence beside the this branch regression. [R4 / packages/skill-evals/src/suites/first-tree-seed/grader.ts:1162, packages/skill-evals/src/suites/first-tree-seed/grader.ts:1170, packages/skill-evals/src/suites/first-tree-seed/grader.ts:1188]

✅ 2. The independent-clause this branch case is now accepted, while bare review this still preserves the candidate PR and the complete-context fixes remain intact.

Action taken

• Submitted request changes.

[Gandy2025]

Fixed in f423b456.

The noun-aware competing-referent matcher now also handles demonstrative this / that / the work|changes phrases, including object-before-agent wording. It remains narrow enough not to classify the verb work as a mergeable referent.

Added both reviewer-provided internal-work variants (That work came from the sub-agent and This branch came from the sub-agent) beside the valid bare-review this case. Validation:

• focused seed grader/floor/coverage: 136 passed
• @first-tree/skill-evals typecheck: passed
• Biome on changed files: passed
• git diff --check: passed

[baixiaohang]

Review by baixiaohang-code-reviewer - PR #1768 (follow-up at f423b456)

Recommendation: request changes

• Rationale: Object-before-agent internal work is fixed, but classifying every demonstrative work phrase as competing creates the inverse false negative for ordinary PR review prose.

Risk level: A

• Path baseline: includes packages/client/** -> A
• Semantic lift: none

PR summary

• Author / repo: Gandy2025 / agent-team-foundation/first-tree
• Problem: Context Tree bootstrap can strand teams between a structure PR and a later content pass that depends on a merge-and-ping handoff.
• Approach: Use one human tree-owner confirmation, build structure and initial leaves in one run, and deliver one seed PR; deterministic evals distinguish that lifecycle from the retired staged-PR flow.
• Impacted modules: skills/first-tree-seed, generated client briefing copy, packages/skill-evals, and skills/first-tree-welcome.

Review findings
❌ 1. The new alternative treats every the|this|that work|change phrase as a competing internal referent, even when it plainly refers to the candidate PR's contents. The retired response “I will deliver one PR for structure and leaves. I will open a structure PR for review. You can review this work first. Once merged, I will add the leaves.” now sets the referent to competing at this work, so the passive PR merge is ignored and the mixed one-PR claim passes without legacy_two_pr_handoff. This is ordinary review language for the structure PR, not an internal branch/draft. Please require internal provenance/mergeable-object context before work / changes switches the referent (while keeping the supplied came from the sub-agent cases), and add this sentence as the opposite-direction regression. [R4 / packages/skill-evals/src/suites/first-tree-seed/grader.ts:1162, packages/skill-evals/src/suites/first-tree-seed/grader.ts:1170]

✅ 2. The reported That work came from the sub-agent and This branch came from the sub-agent cases are now accepted, including the original comma-joined form through the candidate-clause scan.

Action taken

• Submitted request changes.

[Gandy2025]

Fixed in 664a02cf.

this|that work|changes now switches to a competing referent only when the clause carries internal provenance such as came from the sub-agent; ordinary PR review prose such as review this work preserves the candidate PR. For overlapping demonstrative phrases, competing noun events have explicit precedence over the bare demonstrative event.

Added the reviewer-provided review this work retired-flow regression and the comma-joined internal-provenance variant. Validation:

• focused seed grader/floor/coverage: 136 passed
• @first-tree/skill-evals typecheck: passed
• Biome on changed files: passed
• git diff --check: passed

[baixiaohang]

Review by baixiaohang-code-reviewer - PR #1768 (follow-up at 664a02cf)

Recommendation: request changes

• Rationale: Provenance now disambiguates this work, but the equivalent definite-reference PR review wording still has no candidate event and leaves a normal retired flow undetected.

Risk level: A

• Path baseline: includes packages/client/** -> A
• Semantic lift: none

PR summary

• Author / repo: Gandy2025 / agent-team-foundation/first-tree
• Problem: Context Tree bootstrap can strand teams between a structure PR and a later content pass that depends on a merge-and-ping handoff.
• Approach: Use one human tree-owner confirmation, build structure and initial leaves in one run, and deliver one seed PR; deterministic evals distinguish that lifecycle from the retired staged-PR flow.
• Impacted modules: skills/first-tree-seed, generated client briefing copy, packages/skill-evals, and skills/first-tree-welcome.

Review findings
❌ 1. The patch correctly restores candidate coreference for ordinary review this work, but candidate pronoun events are still limited to it|this|that; a definite reference to the just-introduced PR contents has no event. Thus “I will deliver one PR for structure and leaves. I will open a structure PR for review. You can review the work first. Once merged, I will add the leaves.” leaves candidatePrCoreferenceObserved=false at passive index 2 and still misses the retired intermediate merge. Here the work follows an explicit structure PR for review and has no internal provenance, so it is the same candidate-content reference as the now-covered this work. Please preserve the candidate through this definite-reference review form while retaining the provenance-qualified competing cases, and add the sentence as a regression. [R4 / packages/skill-evals/src/suites/first-tree-seed/grader.ts:1162, packages/skill-evals/src/suites/first-tree-seed/grader.ts:1170, packages/skill-evals/src/suites/first-tree-seed/grader.ts:1193]

✅ 2. Explicit internal provenance now wins over overlapping demonstrative events, and the comma-joined internal-work plus ordinary review this work cases are correctly distinguished.

Action taken

• Submitted request changes.

[yuezengwu]

Review at 664a02cf

Recommendation: approve

Goal and core change: This PR replaces the stranding-prone two-PR seed lifecycle with one explicit human tree-owner domain/owner confirmation, followed by structure and initial leaves in one run and one reviewable seed PR. It updates the shipped skill, client briefing and welcome routing, contributor documentation, and deterministic eval contract consistently around that boundary.

The accumulated lifecycle blockers are resolved. In the final referent fix, provenance-bearing internal work (that work came from the sub-agent) remains distinct from the candidate PR, while ordinary PR review prose (review this work) preserves the candidate structure-PR referent. Both directions, the comma-joined form, and the earlier active/passive/landing/punctuation cases are covered by regression tests.

Risk level: A — packages/client/** is in the change surface and this changes a core onboarding lifecycle. There is no database/schema change or persisted core-data migration. Human behavior review should retain the domain checklist as the sole approval gate and confirm that no intermediate merge/ping handoff remains.

Validation at this head:

• focused seed grader/floor/coverage: 136/136 pass
• @first-tree/skill-evals typecheck: pass
• pnpm validate:skill: pass
• touched client briefing test: 26/26 pass
• pnpm check: pass with the repository's existing two warnings and one info
• git diff --check origin/main...HEAD: pass

The linked draft Context Tree PR is aligned with the confirmation-boundary + single-PR contract and can be reconciled after this implementation merges.

[Gandy2025]

Fixed in 4c2813a1.

Definite candidate-content references (the work / the changes) now preserve the active structure-PR referent. Provenance-qualified internal-work events still run first at the same position, so the work came from the sub-agent remains a competing referent.

Added the exact review the work retired-flow regression. Validation:

• focused seed grader/floor/coverage: 136 passed
• @first-tree/skill-evals typecheck: passed
• Biome on changed files: passed
• git diff --check: passed

[baixiaohang]

Review by baixiaohang-code-reviewer - PR #1768 (follow-up at 4c2813a1)

Recommendation: request changes

• Rationale: Definite candidate content is fixed, but the same referent grammar omits the ordinary plural demonstratives for changes, leaving both directions unclassified.

Risk level: A

• Path baseline: includes packages/client/** -> A
• Semantic lift: none

PR summary

• Author / repo: Gandy2025 / agent-team-foundation/first-tree
• Problem: Context Tree bootstrap can strand teams between a structure PR and a later content pass that depends on a merge-and-ping handoff.
• Approach: Use one human tree-owner confirmation, build structure and initial leaves in one run, and deliver one seed PR; deterministic evals distinguish that lifecycle from the retired staged-PR flow.
• Impacted modules: skills/first-tree-seed, generated client briefing copy, packages/skill-evals, and skills/first-tree-welcome.

Review findings
❌ 1. The added candidate-content event recognizes only the changes|work, while the demonstrative vocabulary elsewhere remains singular this|that; the direct plural form has no referent event. Consequently, “I will deliver one PR for structure and leaves. I will open a structure PR for review. You can review these changes first. Once merged, I will add the leaves.” still misses the retired intermediate merge because these changes does not preserve the candidate through passive index 2. The inverse provenance form, These changes came from the sub-agent, is also absent from the competing pattern. Please include ordinary these|those changes in the same provenance-aware event model and add both candidate-review and internal-provenance regressions. [R4 / packages/skill-evals/src/suites/first-tree-seed/grader.ts:1162, packages/skill-evals/src/suites/first-tree-seed/grader.ts:1170, packages/skill-evals/src/suites/first-tree-seed/grader.ts:1174]

✅ 2. review the work now preserves the candidate PR, while equal-position provenance-qualified internal work still wins through event priority.

Action taken

• Submitted request changes.

[yuezengwu]

Review at 4c2813a1

Recommendation: request changes

Goal and core change: This PR replaces the stranding-prone two-PR seed lifecycle with one human tree-owner checklist confirmation, followed by structure and initial leaves in one run and one reviewable seed PR. The remaining review surface is the deterministic oracle that must reject the retired merge handoff without rejecting compliant same-PR builds.

Risk level: A — packages/client/** is in scope and the PR changes a core onboarding lifecycle. There is no database/schema change or persisted core-data migration.

Blocking finding

❌ The referent grammar at this head handles singular/definite this|that|the work|changes, but omits plural demonstratives these|those changes in both candidate and provenance-qualified internal references. I reproduced both directions by adding these sentences to the existing matrices:

1. Retired flow incorrectly accepted (legacyHandoffObserved=false instead of true):

   I will deliver one PR for structure and leaves. I will open a structure PR for review. You can review these changes first. Once merged, I will add the leaves.

2. Compliant same-PR flow incorrectly rejected (legacyHandoffObserved=true instead of false):

   I will open one seed PR with the structure, and these changes came from the sub-agent. Once merged, I will add the initial leaves to that same PR before you merge it.

Please include these|those changes in the same provenance-aware event model and commit both opposite-direction regressions. [packages/skill-evals/src/suites/first-tree-seed/grader.ts:1162]

The committed baseline remains green (focused seed grader/floor/coverage 136/136; skill-evals typecheck passes), which confirms these cases are absent rather than an existing-suite regression. The temporary probes were removed and the review worktree is clean.

This review supersedes my approval on the prior SHA: the concurrent the work finding arrived after that turn closed, and the current replacement head exposed the plural gap during follow-up verification.

[Gandy2025]

Fixed in c272c64e.

The provenance-aware referent model now includes plural these|those changes in both directions: ordinary review language preserves the candidate structure PR, while explicit internal provenance (came from the sub-agent) makes the changes a competing merge referent. Competing events retain precedence at overlapping positions.

Added candidate-review and internal-provenance regressions for both these and those. Validation:

• focused seed grader/floor/coverage: 136 passed
• @first-tree/skill-evals typecheck: passed
• Biome on changed files: passed
• git diff --check: passed

[yuezengwu]

Review at c272c64e

Recommendation: approve

Goal and core change: This PR replaces the stranding-prone two-PR seed lifecycle with one explicit human tree-owner domain/owner confirmation, followed by structure and initial leaves in one run and one reviewable seed PR. The shipped skill, routing copy, contributor documentation, and deterministic eval contract now align on that lifecycle.

The accumulated oracle blockers are resolved. The final referent model covers singular, definite, and plural candidate-content language while preserving provenance-qualified internal work/changes as a competing merge referent. In particular, review these/those changes correctly detects the retired structure-PR handoff, while these/those changes came from the sub-agent remains a compliant internal merge context; the comma-joined internal-provenance form also passes the candidate-clause scan.

Risk level: A — packages/client/** is in scope and this changes a core onboarding lifecycle. There is no database/schema change or persisted core-data migration. Human behavior review should confirm that the domain checklist is the sole approval gate and that no intermediate merge/ping handoff remains.

Validation at this head:

• focused seed grader/floor/coverage: 136/136 pass
• exact comma-joined plural provenance probe: pass (temporary local test removed)
• @first-tree/skill-evals typecheck: pass
• pnpm validate:skill: pass
• pnpm check: pass with the repository's existing two warnings and one info
• git diff --check origin/main...HEAD: pass

The linked draft Context Tree PR remains the required durable-contract reconciliation after this implementation merges.

[baixiaohang]

Review by baixiaohang-code-reviewer - PR #1768 (final at c272c64e)

Recommendation: approve

• Rationale: The single-PR seed lifecycle is coherent across the shipped behavior, routing/documentation, deterministic contract, and paired durable-tree draft; all accumulated blocking findings are resolved.

Risk level: A

• Path baseline: includes packages/client/** -> A
• Semantic lift: none

PR summary

• Author / repo: Gandy2025 / agent-team-foundation/first-tree
• Problem: Context Tree bootstrap could strand teams after a structure-only PR, requiring an intermediate merge and return-to-chat handoff before initial content existed.
• Approach: Keep one explicit human tree-owner domain/owner confirmation, then build the approved structure and initial leaves in one run and deliver one reviewable seed PR while preserving structure and per-domain commit boundaries.
• Impacted modules: skills/first-tree-seed, generated client briefing copy, packages/skill-evals, packages/skill-evals/README.md, and skills/first-tree-welcome.

Review findings
✅ 1. The domain/owner checklist remains the sole human approval gate; there is no intermediate merge, ping, return-to-chat step, or second final-diff approval.

✅ 2. The deterministic oracle now distinguishes the retired seed-PR handoff from compliant same-PR internal integration across active/passive merge and landing forms, polarity, punctuation, PR referents, pronoun/demonstrative coreference, and provenance-qualified internal work/changes.

✅ 3. Routing copy, contributor documentation, the skill version, and linked draft Context Tree PR #734 align with the same confirmation-boundary plus single-PR lifecycle.

✅ 4. All current GitHub checks pass. Validation details in the PR and follow-up reviews cover the changed seed evals, client briefing tests, typecheck, skill validation, and repository checks.

Action taken

• Approved with explicit authorization from baixiaohang for head c272c64e.