refactor(tools): move skill token measurement to Go

Problem

Pull request #836 merged the Go port of skill token measurement: agentrepo skill-tokens (internal/agentrepo/skilltokens.go:12) over internal/tokens/measure.go:64. It left activation to a follow-up. The skill-tokens task still runs bash tools/skill-tokens.sh (mise.toml:487), so the shell backend selection, manifest partitioning, file discovery, record assembly and budget reporting (tools/skill-tokens.sh:69, tools/skill-tokens.sh:110, tools/skill-tokens.sh:139, tools/skill-tokens.sh:166, tools/skill-tokens.sh:203, tools/skill-tokens.sh:224) are still what the repository runs, on origin/main at 44be762a34eec4a832e1ed608c77ebf188357719. The Go command has no caller.

The remaining work is to activate and verify that command through the task, with parity fixes only where a check finds a difference from the frozen shell.

This is the token-measurement split of #332. Part of #588.

Acceptance criteria

AC1: Default and explicit skill selection, including missing-manifest skips, equals the frozen command's selection.
AC2: Each skill selects the same counting backend, model and encoding under the existing defaults and environment precedence.
AC3: Codex skill measurement remains local and never invokes the Anthropic counting backend.
~~AC4: Manifest, frontmatter, description, body and bundled-file counts equal the frozen command's for identical backend responses.~~ Merged into AC6, whose record bytes contain every count.
~~AC5: Bundled-file selection includes tracked and nonignored untracked text and excludes manifests, token records, ignored files and non-text files as the frozen command does.~~ Merged into AC6, whose record rows list the bundled files.
AC6: Each skill's token-record bytes after an invocation equal the frozen command's for the same input and backend results. That covers manifest, frontmatter, description, body and bundled-file counts, and bundled-file selection: tracked and nonignored untracked text is included; manifests, token records, ignored files and non-text files are excluded.
AC7: Failed counting, discovery, measurement or total calculation returns nonzero, including a producer that prints plausible data before failing.
AC8: The body-budget threshold remains advisory, with the same warning/report transcript and a successful status when all measurements succeed.
~~AC9: Existing entry points end as thin shell wrappers that probe the Go subcommand without side effects and fall back to frozen shell only when that capability is absent.~~ Dropped: the design no longer edits the shell file, and the dispatch moves into the task (AC13, AC14).
~~AC10: A supported Go failure never invokes frozen shell or repeats a counting request.~~ Dropped: the task never reaches frozen shell. The no-repeat half is covered by AC13.
AC11: Non-functional: release N-2 invocations, environment settings and record consumers remain compatible without raising the required binary floor.
AC12: Non-functional: identical input text and backend responses produce the same records on supported Linux, macOS and Windows environments.
AC13: The skill-tokens mise task runs the checkout's agentrepo skill-tokens source (go run ./cmd/agentrepo skill-tokens) with the same arguments. No shell file is on its path, no agentrepo binary found on PATH is trusted, and a failure triggers no repeated counting request.
AC14: tools/skill-tokens.sh and its tests are byte-identical to origin/main, and the change contains no changed .sh or .bash file.
AC15: The shell file's retirement is tracked by chore(skill-tokens): retire the frozen shell fallback #951, which is amended so it deletes tools/skill-tokens.sh and its frozen-comparison fixtures after this pull request merges, with no capability-probe wrapper and no N-2 window.
AC16: Non-functional: the pull request declares its consumer impact (Compatibility: none under Risk, with the reason that no consumer calls the shell file) and passes mise run check, mise run cover-go-check and mise run cover-py-check.

Scope and non-goals

Preserve backend classification for packaged/deployed Codex paths versus other skill paths; model/encoding defaults and overrides; one measured Claude request envelope per invocation; manifest partitioning and folded-description extraction; byte sizes; sorted bundled rows; relative recorded paths; backend-specific encoding metadata; deterministic JSON without timestamps; budget-at-or-above comparisons; skip text, summaries and final guidance. Source evidence: tools/skill-tokens.sh:44, tools/skill-tokens.sh:114, tools/skill-tokens.sh:147, tools/skill-tokens.sh:166, tools/skill-tokens.sh:203, tools/skill-tokens.sh:240 and tools/skill-tokens.sh:266.

The Go engine and its unit tests exist from #836. This change reuses them and adds only the task change, parity and failure fixtures that compare against the frozen shell, and focused corrections for any difference those fixtures find. Port or cover every behavior exercised by tests/scripts/skill-tokens-backends.bats:29 through tests/scripts/skill-tokens-backends.bats:86. Preserve the failure boundary before writing a skill record when that skill's counts, file discovery or byte measurements fail (tests/scripts/skill-tokens-backends.bats:39, tests/scripts/skill-tokens-backends.bats:52, tests/scripts/skill-tokens-backends.bats:65). A later skill's failure does not roll back earlier completed records; the final-total regression asserts failure status, not whole-invocation rollback.

Non-goals: changing tokenization, models, encodings, authentication or provider contracts; adding another counting service; imposing a hard body limit; generalizing frontmatter syntax; making the entire invocation transactional; changing skill generation and deployment; or deleting the shell file (#951).

Approach

Change the skill-tokens task body at mise.toml:487 to go run ./cmd/agentrepo skill-tokens, the way other tasks start Go from the checkout (mise.toml:290 and mise.toml:604). Running the checkout's source means the task uses the reviewed code, never an older agentrepo on PATH, and raises no installed-binary floor (AC11). Don't route through the generic run task: it executes Bash and a shell helper (mise.toml:196, mise.toml:200), which would violate AC13. Leave tools/skill-tokens.sh and its tests untouched: the shell-allowlist gate checks every command of any changed .sh file (cmd/shell-allowlist/changed.go:30), and moving the body into that file failed it.

No consumer calls tools/skill-tokens.sh directly: the only caller is mise.toml:487, and the remaining occurrences are tests, the generated_by provenance string, documentation and comments. tools/AGENTS.md marks tools/ internal. This was checked across the tracked tree only, not external clones. The shell file therefore needs no capability-probe wrapper or N-2 window, and #951 deletes it after this merges.

Start with failed-count and failed-discovery checks, then compare the Go command with frozen shell using deterministic backend responses. Preserve the generation order (mise.toml:466, mise.toml:477). Leave code structure to the pull request.

Superseded pull requests

The new pull request replaces #937 and #987. Both edit the shell file into a wrapper with a legacy fallback, so neither implements this design. The new pull request's description says so, begins a fresh review count, and does not inherit approvals or findings from either. After it opens, #937 and #987 are closed as superseded, each with a comment naming the new pull request. Neither keeps a Closes #810 reference.

Dependencies

Native blocked-by edges on this issue: none to set. The live blockedBy list is empty, and that is correct.

feat(tokens): add deterministic skill measurement #836 delivered the Go command and is merged.
refactor(release): remove obsolete dispatch orchestration #542 is dropped as a dependency. It removes obsolete release orchestration, and this change touches no release interface. The prose line and the delivery chain through Ship the Go binaries as release assets installed through mise's github backend #333, feat(distribution): migrate consumer tool pins to release assets #544, feat(release): provide configurable shared release code and tasks #382, feat(release): distribute the reusable workflow and release skills #545, build(release): adopt the shared release path in the first consumer #546 and Release on every merge that has no release dependencies #359 are removed with it.
Blocker to set on chore(skill-tokens): retire the frozen shell fallback #951: refactor(tools): move skill token measurement to Go #810, as a native blocked-by edge, and its text is amended per AC15. Its current "blocked by the N-2 window" prose and wrapper requirement no longer apply.

#332 and #588 track this split. No unpublished sibling draft is needed to begin.

Size

M. One pull request for task activation, parity and failure fixtures, and any focused corrections. It reuses the #836 implementation and stays under 800 code lines.

Verify plan

AC7: first reproduce each failure case at tests/scripts/skill-tokens-backends.bats:39, tests/scripts/skill-tokens-backends.bats:52, tests/scripts/skill-tokens-backends.bats:65, tests/scripts/skill-tokens-backends.bats:74 and tests/scripts/skill-tokens-backends.bats:86. Cover silent failure, plausible output followed by failure and invalid counts; assert status and the existing per-record preservation boundary, without imposing invocation-wide rollback.
AC1: compare default discovery, explicit directories, trailing separators and absent manifests with frozen shell.
AC2, AC3: backend/model/encoding fixtures cover default values, every override and precedence combination, relative/absolute input spellings, packaged and deployed Codex paths and non-Codex paths. An Anthropic invocation trap remains untouched for Codex, extending tests/scripts/skill-tokens-backends.bats:29.
AC6 (counts): deterministic responses cover manifest sections, folded descriptions, missing frontmatter, empty and UTF-8 text, multiple Claude skills and envelope subtraction; compare each count and the single baseline request.
AC6 (selection): tracked, untracked, ignored, empty, binary and text fixtures compare the sorted bundled set, skip diagnostics and byte sizes.
AC6: compare exact record bytes, keys, optional encoding metadata, relative paths and totals; repeat unchanged input and assert no content drift. A two-skill failure fixture confirms the first completed record remains.
AC8: counts just below, equal to and above the budget compare warnings, per-skill summaries, final guidance and successful status.
AC13: read the task definition and run it against a stub backend with an older stand-in agentrepo first on PATH; assert the checkout's source runs, the arguments arrive, no .sh file runs, and a failure after partial output fails the task with one recorded backend request.
AC14: git diff --stat origin/main lists no .sh or .bash path, and mise run lint-shell-allowlist passes.
AC15: read chore(skill-tokens): retire the frozen shell fallback #951 after amendment: its body has no wrapper or probe requirement, names the deleted paths, and the live blockedBy of chore(skill-tokens): retire the frozen shell fallback #951 lists refactor(tools): move skill token measurement to Go #810.
AC16: the description carries the compatibility line under Risk; mise run check, mise run cover-go-check and mise run cover-py-check pass on the reviewed head.
AC11: run the changed task and token-record readers against N-2 record fixtures.
AC12: run the Go corpus on Linux, macOS and Windows. All verification uses local deterministic backend stubs, without real API counting or authentication changes. Run focused tests, generation/deployment consistency checks, mise run lint and mise run test on the reviewed head.
Definition of done: D1 each AC is verified at the merged head, with the commands and results cited in the pull request's Verification. D2 an independent review approves the exact merged head, and every finding has a recorded disposition. D3 required checks pass with no gate loosened. D4 the leak scans the repository specifies are clean, and every post made for this work (comments, descriptions) passes the per-post leak review before it is posted. D5 every commit is signed. D6 the documentation matches, including the tools/skill-tokens.sh row in docs/script-inventory.md, or the description says No doc impact with the reason. D7 every commit carries an Assisted-by trailer and the pull request carries the provenance:agent-assisted label. D8 the change stays within the pull request size limits. D9 follow-up work, including deferred review findings, is filed as issues, each with a Problem section and numbered acceptance criteria and linking this issue; the pull request's Related section lists them, including chore(skill-tokens): retire the frozen shell fallback #951 as amended, and any code that defers work links its issue. D10 the merge closes this issue through a closing reference, the project board shows the state Done, and a final Status: comment names the merge commit.

Risks

An incorrect envelope or backend selection corrupts committed token records; nondeterministic metadata creates persistent drift. A task left on the old shell path would leave the port unused. go run compiles on each invocation, which adds startup time to regen-generated. Exact record comparisons and invocation-count assertions cover the first boundaries. Undo: revert the task change.

Operator authority

None.

Refinement: READY — skill token measurement port to Go has clear, testable requirements and verified source boundaries
Model: gemini-4-argon-high

Checked

Live issue and referenced issues: Read live tbhb-dev/devtools#810 (0 comments) and referenced issues #332, #588, #542, #333, #544, #382, #545, #546, and #359.
Baseline and Problem boundaries against origin/main (a0347a715fab580ece45edfcb3d13509620459a8):
- tools/skill-tokens.sh:69 (count_claude backend invocation; count_codex follows at tools/skill-tokens.sh:87).
- tools/skill-tokens.sh:110 (prepare_backend single-request Claude envelope baseline measurement at tools/skill-tokens.sh:110-120).
- tools/skill-tokens.sh:139 (skill directory loop for dir in "${dirs[@]}", trailing slash normalization, missing SKILL.md skip to stderr, and path-based backend classification at tools/skill-tokens.sh:139-158).
- tools/skill-tokens.sh:166 (embedded Python manifest partitioning into frontmatter, body, and folded description at tools/skill-tokens.sh:166-177).
- tools/skill-tokens.sh:203 (git -c core.quotePath=false ls-files --cached --others --exclude-standard -- "$dir" | sort discovery, SKILL.md/tokens.json exclusion, grep -Iq . non-text skip, per-file net count, and byte measurement at tools/skill-tokens.sh:203-219).
- tools/skill-tokens.sh:224 (embedded Python tokens.json assembly with relative paths, sorted keys, 2-space indentation, optional tokenizer, and no timestamp at tools/skill-tokens.sh:224-264, followed by advisory body-budget warning and summary output at tools/skill-tokens.sh:266-284).
Acceptance criteria (AC1–AC12):
- AC1–AC10 are observable, testable, single-outcome behavioral requirements free of prescribed internal Go package layout or implementation details.
- AC11 and AC12 state the non-functional requirements for release N-2 compatibility (docs/compatibility.md:1-5) and cross-platform record determinism across Linux, macOS, and Windows.
Scope, non-goals, and Approach:
- Source citations match origin/main: tools/skill-tokens.sh:44 (BODY_BUDGET default and model/encoding environment variables at tools/skill-tokens.sh:44-49), tools/skill-tokens.sh:114 (baseline measurement), tools/skill-tokens.sh:147 (Codex vs. Claude path matching), tools/skill-tokens.sh:166 (manifest split), tools/skill-tokens.sh:203 (bundled file discovery), tools/skill-tokens.sh:240 (timestamp-free JSON record structure), and tools/skill-tokens.sh:266 (-ge budget check and summary formatting).
- Existing Bats coverage citations match origin/main: tests/scripts/skill-tokens-backends.bats:29 (Codex local counting without touching Anthropic trap), tests/scripts/skill-tokens-backends.bats:39 (bundled tokenizer failure writes no record), tests/scripts/skill-tokens-backends.bats:52 (git ls-files discovery failure writes no record), tests/scripts/skill-tokens-backends.bats:65 (wc byte measurement failure writes no record), tests/scripts/skill-tokens-backends.bats:74 (Claude count failure propagation through net), and tests/scripts/skill-tokens-backends.bats:86 (failed bundled total calculation returns nonzero after record write).
- Task and generation-order citations match mise.toml:466 ([tasks.regen-generated], tested in tests/tasks/regen-generated.bats:35-42) and mise.toml:477 ([tasks.skill-tokens]).
- Design and internal code structure are explicitly left to the pull request.
Dependencies:
- Named issues #542, #333, #544, #382, #545, #546, #359, #332, and #588 exist and are open; prerequisite (#542 following the #333–#546 and #359 chain) and tracking epics (#332, #588) are accurately classified.
Size:
- Size M fits a single pull request covering the Go measurement port, record compatibility, failure coverage, and thin entry point wrapper.
Verify plan:
- Maps every acceptance criterion (AC1–AC12) and D1–D10 to concrete, offline-capable verification checks using deterministic backend stubs.
Risks and undo:
- Identifies record corruption, nondeterministic drift, and duplicate backend calls on fallback, paired with mitigations and a clear undo path (revert the port).
Operator authority:
- None. Verified against .agents/agent-skills.conf:4-9 (golden_paths, constitutional_paths, gate_config_paths, owner_paths) and .github/CODEOWNERS:1-13: porting tools/skill-tokens.sh to Go while keeping mise.toml:466 and mise.toml:477 unchanged does not touch protected owner/golden paths or loosen any gate.

Gaps

None.

Split

None needed; Size M fits in one pull request.

Definition of done

The definition of done for #810 (Feature). Tick each item and give its evidence. All ticked boxes below indicate planning coverage in the issue's acceptance criteria and verify plan, not completed implementation work.

Round: 1 of 3

feat(tools): port skill token measurement to Go

fix(tokens): preserve frozen discovery and summary output

feat(tokens): add deterministic skill measurement

refactor(tokens): use Go measurement from the shell entry point

Refinement: NOT READY — option 2 is viable, but dependencies and remaining scope need reconciliation

Checked

Read refactor(tools): move skill token measurement to Go #810 and its one earlier refinement comment, native parent Port the tools/ and release scripts #332, siblings refactor(tools): move skill generation to Go #807–refactor(tools): move script hygiene checks to Go #809 and refactor(release): move readiness diagnostics to Go #811–refactor(tools): move mise bootstrap logic to Go #813, governing Epic Replace complex shell with tested Go, starting with gates and merging #588, prerequisite refactor(release): remove obsolete dispatch orchestration #542, follow-up chore(skill-tokens): retire the frozen shell fallback #951, merged PR feat(tokens): add deterministic skill measurement #836 and open PRs refactor(tokens): use Go measurement from the shell entry point #937/refactor(tokens): use Go measurement from the shell entry point #987. This is refinement round 2; a fresh PR review count does not reset the issue's refinement count.
Independently matched origin/main to live remote main at 44be762a34eec4a832e1ed608c77ebf188357719.
Option 2 can avoid new shell logic: the production caller is mise.toml:487; direct Go task precedents are mise.toml:290 and mise.toml:604. The allowlist selects changed shell files at cmd/shell-allowlist/changed.go:30. That filter does not authorize shell logic inside TOML. The generic run task executes Bash and a shell helper at mise.toml:196 and mise.toml:200; routing through it would violate AC13. Direct Go dispatch avoids that conflict.
A tracked-tree reference search found no published consumer calling tools/skill-tokens.sh: the production call is mise.toml:487; other occurrences are tests, provenance, documentation or comments. tools/AGENTS.md:3 identifies these tools as internal. This supports the body's answer within this repository, without claiming knowledge of arbitrary external clones.
Read the frozen backend, envelope, selection, partition, discovery, record and reporting boundaries at tools/skill-tokens.sh:69, tools/skill-tokens.sh:110, tools/skill-tokens.sh:139, tools/skill-tokens.sh:166, tools/skill-tokens.sh:203, tools/skill-tokens.sh:224 and tools/skill-tokens.sh:266, and the failure cases at tests/scripts/skill-tokens-backends.bats:39, tests/scripts/skill-tokens-backends.bats:52, tests/scripts/skill-tokens-backends.bats:65, tests/scripts/skill-tokens-backends.bats:74 and tests/scripts/skill-tokens-backends.bats:86.
Every active AC (AC1–AC8, AC11–AC14) has a Verify bullet; AC9/AC10 are explicitly retired. Compatibility, deterministic records, local Codex counting, partial-output failure and per-record preservation are testable. AC13/AC14 deliberately constrain dispatch to support the freeze. Risks and task-only rollback are stated. M is a plausible ceiling below 800 code lines when reusing the delivered implementation.
The Approach says the replacement PR replaces refactor(tokens): use Go measurement from the shell entry point #937 and refactor(tokens): use Go measurement from the shell entry point #987 and starts a fresh review count. Carry that statement into the actual PR description. Both earlier PRs remain open and neither implements option 2.
No mechanical gate was rerun. No test suite or counting API was run.

Gaps

Dependencies are not native and appear stale. The body says Depends on refactor(release): remove obsolete dispatch orchestration #542, which remains open, but the native blocked-by endpoint returns zero dependencies. refactor(release): remove obsolete dispatch orchestration #542 concerns obsolete release orchestration; sibling tooling issues refactor(tools): move skill generation to Go #807/refactor(tools): move script hygiene checks to Go #809 explicitly need no such release-chain dependency. Either remove this obsolete dependency with a rationale or retain it, explain the required handoff and add the native blocked-by edge. Do not leave a prose-only blocker.
Describe the remaining work after merged PR feat(tokens): add deterministic skill measurement #836. Go measurement and its command already exist at internal/tokens/measure.go:64 and internal/agentrepo/skilltokens.go:12. feat(tokens): add deterministic skill measurement #836 explicitly left caller activation to a follow-up. Reframe Problem/Approach/Size around activating and verifying that command, with focused parity fixes only as needed. State how the task uses the checkout's command without raising the installed binary floor (AC11), executing shell files (AC13) or trusting an older PATH binary. Direct Go task precedents provide a viable route. Require the replacement statement and fresh review count in the new PR text, and explicitly disposition the two open superseded PRs.
Align completion and cleanup with option 2. Existing follow-up chore(skill-tokens): retire the frozen shell fallback #951 still requires an N-2 capability-probe wrapper, contrary to this body's conclusion that the internal shell path needs no consumer window. Name and reconcile that follow-up under D9 so cleanup does not recreate the rejected wrapper. Add the repository-specific completion checks: mise run check, mise run cover-go-check and mise run cover-py-check; lint/test alone do not cover them. Include the task's consumer-impact declaration or the reason for no consumer impact.

Split

No further implementation split is needed. #836 already delivered the Go engine. Task activation, argument/failure integration checks and bounded parity corrections fit one PR below 800 code lines. Shell deletion remains a separately tracked follow-up. Retain #588's freeze and do not edit shell to implement dispatch.

Definition of done

Checked boxes indicate explicit planning coverage in the ACs/Verify plan, not completed implementation. The gaps above still prevent readiness.

Round: 2 of 3

Refinement: READY — activate the delivered Go command through the task with frozen-shell parity checks

Checked

Read the revised refactor(tools): move skill token measurement to Go #810 body and its two earlier Refinement comments: round 3. Read native parent Epic Port the tools/ and release scripts #332, siblings refactor(tools): move skill generation to Go #807–refactor(tools): move script hygiene checks to Go #809 and refactor(release): move readiness diagnostics to Go #811–refactor(tools): move mise bootstrap logic to Go #813, governing Epic Replace complex shell with tested Go, starting with gates and merging #588, follow-up chore(skill-tokens): retire the frozen shell fallback #951, merged PR feat(tokens): add deterministic skill measurement #836 and open PRs refactor(tokens): use Go measurement from the shell entry point #937/refactor(tokens): use Go measurement from the shell entry point #987. The issue search also finds closed, unmerged refactor(tokens): use Go measurement from the shell entry point #979.
Confirmed origin/main equals live remote main at 44be762. The production caller remains mise.toml:487. The delivered command at internal/agentrepo/skilltokens.go:12 delegates to internal/tokens/measure.go:64; feat(tokens): add deterministic skill measurement #836 explicitly left activation undone. Neither open wrapper PR implements the revised direct-task design. The remaining work is not already delivered.
Read the frozen backend/envelope, selection, partition, discovery, record and reporting boundaries at tools/skill-tokens.sh:69, tools/skill-tokens.sh:110, tools/skill-tokens.sh:139, tools/skill-tokens.sh:166, tools/skill-tokens.sh:203, tools/skill-tokens.sh:224 and tools/skill-tokens.sh:266, alongside the Go implementation. Failure witnesses remain at tests/scripts/skill-tokens-backends.bats:39, tests/scripts/skill-tokens-backends.bats:52, tests/scripts/skill-tokens-backends.bats:65, tests/scripts/skill-tokens-backends.bats:74 and tests/scripts/skill-tokens-backends.bats:86.
All 12 active criteria (AC1, AC2, AC3, AC6, AC7, AC8, AC11–AC16) have explicit Verify bullets. Retired AC4/AC5/AC9/AC10 state their dispositions. Selection, exact records, local Codex counting, failure boundaries and advisory reporting are testable. Compatibility, platform determinism and completion gates are explicit non-functional requirements. AC13/AC14 constrain the dispatch boundary to enforce the shell freeze; internal Go structure remains unconstrained.
The Approach reuses feat(tokens): add deterministic skill measurement #836 and directly runs the checkout's source, following mise.toml:290 and mise.toml:604. It excludes the generic shell-backed task at mise.toml:196 and mise.toml:200. No new shell dispatch, probe or fallback is proposed. The shell and its existing tests remain frozen; retirement is separate under chore(skill-tokens): retire the frozen shell fallback #951.
The live native blocked-by list on refactor(tools): move skill token measurement to Go #810 is empty, consistent with removing the unrelated release-chain dependency. Unlike release siblings refactor(release): move readiness diagnostics to Go #811/refactor(release): move release verification to Go #812, this activation needs no completed release diagnostic handoff or installed replacement binary.
Read chore(skill-tokens): retire the frozen shell fallback #951: its body still describes an N-2 wrapper and its native blocked-by list is empty. AC15 and its Verify bullet explicitly require replacing that stale design with shell/fixture deletion after refactor(tools): move skill token measurement to Go #810 and setting chore(skill-tokens): retire the frozen shell fallback #951 blocked by refactor(tools): move skill token measurement to Go #810. This is a named completion deliverable, not an unstated prerequisite to starting refactor(tools): move skill token measurement to Go #810. Verify it before closing this issue.
A tracked-tree caller search supports the internal-tool conclusion at tools/AGENTS.md:3: other references are tests, provenance, documentation or comments. This supports the compatibility rationale without claiming knowledge of external clones. The N-2 requirement is preserved without raising the installed binary floor (docs/compatibility.md:3).
Risks cover record corruption, drift, unused activation and compile startup cost; undo reverts the task change. Operator authority None is consistent with the bounded task/measurement change: mise.toml is gate configuration (.agents/agent-skills.conf:8), so normal policy checks still apply, but this issue proposes no gate loosening or protected-path policy edit.
The replacement PR must name refactor(tokens): use Go measurement from the shell entry point #937/refactor(tokens): use Go measurement from the shell entry point #987 as superseded, begin fresh PR review, and close both with replacement links and removal of their closing references, as the issue requires. Existing approvals do not transfer.
Accepted the supplied mechanical-gate pass; did not rerun it. No tests, counting APIs, full suites or skills-package tests were run for this refinement.

Gaps

None blocking. Round 2's dependency, remaining-scope and cleanup/completion gaps are addressed in the revised body. The #951 amendment/native edge and superseded-PR dispositions remain implementation completion checks, not claims of actions already performed.

Split

No further split needed. #836 already delivered the engine and command. Task activation, deterministic parity/failure fixtures and focused corrections fit one M pull request under 800 code lines. Shell deletion remains #951; no shell logic is added here.

Definition of done

Ticks indicate independently checked planning coverage in the current ACs and Verify plan, not completed implementation.

Repository additions are also planned: AC16 requires check and both coverage floors; the Approach preserves generation/token/deployment order; AC16 supplies the compatibility declaration and no-consumer-call rationale. Carry the consumer-impact explanation into PR Verification as well as Risk.

Round: 3 of 3

feat(tools): run skill-tokens from the Go command

Navigation Menu

Repository navigation

refactor(tools): move skill token measurement to Go #810

Problem

Acceptance criteria

Scope and non-goals

Approach

Superseded pull requests

Dependencies

Size

Verify plan

Risks

Operator authority

Checked

Gaps

Split

Definition of done

Checked

Gaps

Split

Definition of done

Checked

Gaps

Split

Definition of done

Metadata

Assignees

Labels

Type

Fields

Projects

Milestone

Relationships

Development

refactor(tools): move skill token measurement to Go #810

Description

Problem

Acceptance criteria

Scope and non-goals

Approach

Superseded pull requests

Dependencies

Size

Verify plan

Risks

Operator authority

Activity

tbhb-agent commented on Oct 26, 2026

Checked

Gaps

Split

Definition of done

tbhb-dev-reviews commented on Oct 31, 2026

Checked

Gaps

Split

Definition of done

tbhb-dev-reviews commented on Oct 31, 2026

Checked

Gaps

Split

Definition of done

Metadata

Metadata

Assignees

Labels

Type

Fields

Projects

Milestone

Relationships

Development

Issue actions