[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7AGWS1G7F5W38K02FPNJ23D":3,"skill-same:01M7AGWS1G7F5W38K02FPNJ23D":51},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":24,"commitSha":25,"skillDirectory":26,"forksCount":27,"repositoryHtmlUrl":28,"repositoryHomepageUrl":6,"ownerHtmlUrl":29,"ownerBlogUrl":6,"ownerDescription":6,"updatedAt":30,"install":31,"resources":44},"01M7AGWS1G7F5W38K02FPNJ23D","sb-dev-forward-test",null,"Run the behavioral forward tests for SpecBind's embedded skills against a fixture project, using subagents with no prior context. Use when a product-managed skill changed materially, before a release, or when a decision's effect on agent behavior needs checking.","Huruikagi\u002Fspecbind","Huruikagi","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F10933612?v=4","01M7AGWP8QKD1EXMV26YA1ESXF","ふるいかぎ","01M7AGWP9WBJ1EEJNQKWB77GN3",2,"MIT","https:\u002F\u002Fgithub.com\u002FHuruikagi\u002Fspecbind\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fsb-dev-forward-test\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: sb-dev-forward-test\ndescription: Run the behavioral forward tests for SpecBind's embedded skills against a fixture project, using subagents with no prior context. Use when a product-managed skill changed materially, before a release, or when a decision's effect on agent behavior needs checking.\n---\n\n# Run the skill forward tests\n\nThe [forward-test index](..\u002F..\u002F..\u002Fdocs\u002Fskill-forward-tests.md) routes to the\nscenario contracts, measurement dashboard, run archive, and findings worklist.\nThe scenario documents are the contract; this skill is how to execute them\nwithout rediscovering the operational traps each time.\n\nThe index also routes to the deliberately expensive end-to-end journey. Use\nthat journey only for cross-lifecycle changes or release-candidate confidence;\ndo not add it to the ordinary per-skill batch.\n\nThis tests **product-managed skills** — the ones under\n`tools\u002Fspecbind\u002Fassets\u002Fskills\u002F` that get installed into consumer projects. It is\nnot about this repository's own workflow.\n\n## Before anything\n\nRecord the commit under test:\n\n```sh\ngit rev-parse --short HEAD\n```\n\nA run measures one build. If you fix something mid-run, every scenario still in\nflight is measuring the old build, and its result is about that build only. Say\nso rather than folding it into the same table.\n\n## One fixture per scenario\n\n```sh\nsh tools\u002Fspecbind\u002Fscripts\u002Fforward-test-fixture.sh \u002Ftmp\u002Fsb-\u003Cscenario> en\n```\n\nNever share a fixture between scenarios. Several depend on the starting state,\nand a leftover milestone silently changes what is being tested. The builder\nrefuses to touch an existing directory, so delete and rebuild.\n\nThe builder prints an `export PATH=` line for the CLI it ships inside the\nfixture. That line is not optional. The skills invoke `specbind` as a bare\ncommand because a real installed project has it on PATH; without it the run stops\nat \"command not found\" and produces no result about the skill at all.\n\nMost scenarios have a recipe that builds their starting state and proves it:\n\n```sh\nsh tools\u002Fspecbind\u002Fscripts\u002Fforward-test-scenario.sh \u003Cscenario> \u002Ftmp\u002Fsb-\u003Cscenario> en\n```\n\nIt wraps the fixture builder, applies the scenario's precondition, and exits\nnonzero with a message when the precondition did not take. Use it rather than\ncomposing the state by hand; that is where a run silently becomes a different\nrun. Add a recipe when a scenario needs one, and give it a check that fails when\nthe setup is a no-op.\n\nThe end-to-end journey has its own prepare-and-judge harness because it starts\nnear the beginning of the lifecycle and its final state crosses several\nscenario families:\n\n```sh\nsh tools\u002Fspecbind\u002Fscripts\u002Fforward-test-journey.sh prepare hp1 \u002Ftmp\u002Fsb-hp1 en\n# Drive the exact conversation in docs\u002Fskill-forward-tests\u002Fjourney-scenarios.md.\nsh tools\u002Fspecbind\u002Fscripts\u002Fforward-test-journey.sh judge hp1 \u002Ftmp\u002Fsb-hp1\n```\n\nDo not run `judge` after repairing the fixture by hand. A failed judgment is the\nmeasurement. Rebuild at a new target and rerun only when a fresh measurement is\nactually worth its agent cost.\n\nScenarios with no recipe yet need a precondition set up first — an uncommitted edit, a\nbroken steering document. **Do that from the shell**, not from a helper script in\nanother language. On Windows the shell maps `\u002Ftmp` to a real directory, but a\nnative interpreter invoked from it does not, so a script that opens\n`\u002Ftmp\u002Fsb-\u003Cscenario>\u002F...` fails while every shell command beside it succeeds. A\nprecondition that silently did not apply turns a scenario into a different one.\nVerify the precondition with a command before launching the run.\n\nAdd `--instrument-dispatch` as the final builder or recipe argument only when\nthe run is intended to prove fresh-context dispatch. It injects the ignored\n`.forward-test\u002Fagents.log` instruction used to distinguish dispatch from the\nmain-context fallback. Ordinary workflow scenarios leave it off: they have no\ndispatch claim to prove, and an unrelated instrumentation write must not block\nthe product workflow before it starts.\n\n## Driving a run\n\nPrefer a real top-level session through the harness whenever the scenario\ncrosses an approval, dispatches, or measures Skill selection. A subagent cannot\ndo any of those faithfully (ENV-0001, ENV-0003):\n\n```sh\nsh tools\u002Fspecbind\u002Fscripts\u002Fforward-test-drive.sh start \u003Cclaude-code|codex> \u002Ftmp\u002Fsb-\u003Cscenario> \"\u003Cquoted request>\"\nsh tools\u002Fspecbind\u002Fscripts\u002Fforward-test-drive.sh send \u002Ftmp\u002Fsb-\u003Cscenario> \"\u003Cconfirmation>\" --expect '\u003Cboundary check>'\n```\n\nThe harness runs the agent's own CLI in the fixture, removes host-session\nvariables, keeps the maintainer's user-level instructions out, and resumes the\nsame session for each turn. Send only the scenario's quoted request: the\nworking-directory, PATH, and stands-alone statements below are what the harness\nalready makes true. Give every confirmation an `--expect` proving the run\nstopped at that boundary; a failed check sends nothing. Exit 5 is\nenvironment-invalid. Read `\u003Cfixture>.drive\u002Fturns\u002FNN-actions.txt` for which Skill\nthe registry selected, and see the forward-test running guide for the full\ncontract.\n\nOtherwise use a subagent with no prior context.\n\n### Codex driver profile\n\nWhen Codex drives a scenario through a subagent, use this profile:\n\n- `fork_turns: \"none\"`\n- `model: \"gpt-6.1-sol\"`\n- `reasoning_effort: \"medium\"`\n\nThis is the default forward-test driver profile, not part of the product skill\nunder test. Override it only when the run deliberately compares models, and\nrecord the override with the result.\n\nA Codex-spawned subagent is a Codex result; a Claude Code Agent-tool subagent is\na Claude Code result. Record the actual driver because the scenario document\ntracks them as separate columns.\n\nThe prompt gives three things and nothing else:\n\n1. the working directory, given as the **native path** its file tools resolve,\n   with the shell alias named separately if the two differ. On Windows a subagent\n   writing to `\u002Ftmp\u002Fsb-\u003Cscenario>\u002F...` with a file tool can land under the drive\n   root instead of the shell's temp mapping, creating a file the CLI then reports\n   as missing. One run caught that itself and moved the file; the next may not.\n2. the `export PATH=` line, stated as an environment fact\n3. the maintainer's request, phrased the way a maintainer would phrase it\n\nThen ask only what it changed and what it ran.\n\nThe fixture's own `AGENTS.md` or `CLAUDE.md` is applicable project context. Say\nthat explicitly when the driver would otherwise treat a prompted working\ndirectory as less authoritative than the host session. If it never reads the\nfixture instructions and installed skill tree, the product skill was not\nmeasured.\n\nOn Windows, also state the native Git Bash `sh.exe` location when a fixture's\ncanonical command uses `sh`. The printed `export PATH=` line still remains the\nenvironment fact for the CLI. A run that cannot resolve either executable is an\nenvironment failure, not a verdict about the product skill.\n\n**Neutralize this repository's instructions.** A subagent is not a clean room: it\ncarries the host session's project instructions (`AGENTS.md` under Codex or\n`CLAUDE.md` under Claude Code), not the fixture's. This repository's own\ninstructions tell an agent to answer in Japanese and to commit each finished unit\nto `main` — and the checkpoint scenarios exist to measure whether the agent\ncommits. State in the prompt that the fixture is a standalone project and that\ninstructions from any other repository do not apply to it. That is environment\nhygiene, not method.\n\nA run reporting in Japanese against an `en` fixture is the visible symptom. Treat\nit as a signal that the contamination is active, and re-read any checkpoint\nresult taken from the same batch.\n\n**Never name a skill or a command in the prompt.** Whether the agent finds and\nuses the installed skill is the thing under test; telling it teaches the answer.\n\n**Never ask it to justify its classification.** An expectation about what the\nagent told the user cannot be measured from a report you asked for — you get the\nsentence because you requested it. Read that from what the run said on its own.\n\n## Expect a confirmation turn\n\nEvery scenario that crosses a guarded transition has one. Discovery confirms\nscope before its mutation; authoring phases may write a draft first and then\nconfirm its approval. A first pass that stops at the applicable boundary is the\nskill working. Judge the draft and guarded state appropriate to that phase\ninstead of assuming the whole repository must still be unchanged.\n\nAnswer as the maintainer would, then let it finish. Continue the same subagent\nrather than starting a new one, so it keeps the state it gathered. The\nconfirm-then-mutate path is part of what the scenario tests, and it is where the\nguarded CLI operations actually run.\n\n**Confirm the phase, not the project.** Refer explicitly to what the agent just\npresented and where to stop: \"I approve the plan you just presented for Discovery\nonly. Stop after Discovery.\" or \"I approve the Requirements and active Requirement\nID selection you just presented. Stop after Requirements.\" If an authoring phase\ndid not present every value its approval accepts, do not supply or infer the missing\nvalue for it; the scenario has not reached an approvable boundary. A bare \"Proceed\nwith that scope\" can read as permission\nto build the whole thing: one run answered that way went from discovery through\nrequirements, design, review, tasks, and implementation, which left every\ndiscovery expectation unmeasurable because later phases had legitimately changed\nthe same files.\n\n## Judge from the fixture, never from the report\n\nRead the result with commands against the fixture:\n\n```sh\nspecbind milestone status\nspecbind milestone scope\nspecbind spec list\nspecbind spec status \u003Cspec>\ngit status --short\n```\n\nA subagent's summary is a claim about what it did. It is usually accurate and it\nis not evidence. Every expectation in the scenario document is readable from the\nfixture; read it there.\n\n## Ask for a usability debrief after judgment\n\nOnly after the fixture has been judged and the pass or failure recorded, ask the\nsame driver for a qualitative usability debrief. Asking earlier contaminates the\nscenario: it teaches the agent which surfaces the maintainer cares about and can\nturn a naturally discovered ambiguity into a prompted answer.\n\nRecord `git status --short` before the debrief. Then continue the same driver\nwith this read-only follow-up; do not start a fresh agent, because the useful\nevidence is what this agent remembers hesitating over:\n\n```text\nThe scenario result has already been judged. Do not run commands or change any\nfiles. Reflect only on the work you just completed.\n\nWhat made you hesitate, infer an unstated fact, take an unnecessary extra step,\nor risk choosing the wrong action? Consider the CLI, skill, template, protocol,\nadapter, diagnostics, and any other product surface that actually affected the\nwork. Do not invent an issue for every category. If there was no friction, say\n\"none\".\n\nFor each observation, report only:\n- Surface: CLI | Skill | Template | Protocol | Adapter | Other\n- Friction: what was difficult or ambiguous\n- Evidence: the exact command, output, wording, or path you encountered\n- Workaround: what you inferred or did to continue\n- Impact: cosmetic | extra-step | ambiguity | wrong-action-risk\n```\n\nAfter the answer, run `git status --short` again. A changed fixture means the\ndebrief violated its read-only boundary; discard the observation, classify the\nmutation before cleanup, and do not fold it into the scenario result.\n\nThe debrief is not pass\u002Ffail evidence and does not overturn a result already\nmeasured from the fixture. It is an exploratory observation. Record it\nseparately, reproduce it against the owning contract before changing the\nproduct, and look for recurrence across scenarios or agents. One\n`wrong-action-risk` is enough to investigate before the next batch; lower-impact\nobservations normally need either concrete reproduction or a repeated pattern.\n\nTreat `docs\u002Fskill-forward-tests\u002Ffindings.md` as a triaged worklist, not an\nappend-only transcript. Keep reproduced unresolved product findings, compact\nresolved findings to their behavior change and fixing build, and retain\nenvironment limitations only while they affect interpretation. Remove\nduplicates, one-off non-defects, fixture-only workarounds, and `none`\nobservations from the worklist after the run itself records their disposition.\n\n## When something fails\n\nDecide which of these it is before changing anything:\n\n- **The skill is wrong.** Fix the skill.\n- **The decision is wrong.** Fix the decision first, then the skill. This is the\n  common case when a scenario and an accepted decision disagree, and it is why\n  the scenario document says to check the decision before editing a row.\n- **The scenario is wrong.** Fix it, but only after confirming the decisions do\n  not require what it asked for.\n- **The harness is wrong.** Fix the fixture builder or this skill, and note that\n  the scenario never actually ran.\n\nNever adjust an expectation until it passes. A forward test tuned to pass has\nstopped measuring anything.\n\nA scenario that fails once and passes on retry is a finding, not a flake: the\nskill is ambiguous enough for the agent to go either way, and that ambiguity is\nthe defect.\n\nThe same applies across agents. Both are handed the identical skill body —\nrendering only rewrites Front Matter — so a scenario that passes under one and\nfails under the other proves the document admits two readings. That is a skill\ndefect, and \"the other agent behaves differently\" restates it rather than\nexplains it.\n\n## Finishing\n\nCopy `docs\u002Fskill-forward-tests\u002Frun-template.md` to\n`docs\u002Fskill-forward-tests\u002Fruns\u002FYYYY-MM-DD-\u003Cdriver>-\u003Cshort-build>[-N].md`. One\nrecord contains one driver and one tested build. Use `pass`,\n`product_failure`, `scenario_invalid`, `environment_invalid`, or\n`environment_blocked` per scenario, and record the expectation that did not\nhold, the state left behind, and fixture evidence. A retry never replaces its\nfailed attempt.\n\nUpdate `docs\u002Fskill-forward-tests\u002Fresults.md` only with the current projection.\nGive reproduced product findings stable `FT-NNNN` identities in\n`docs\u002Fskill-forward-tests\u002Ffindings.md`; use separate `ENV-NNNN` identities for\nactive environment limitations. Record non-findings and discarded observations\nonly in the run record. Then clean up the fixtures.\n\nRecord any defect the run exposed against the decision or skill it belongs to,\nand update the scenario document when the run showed the procedure itself was\nwrong. Both kinds of finding are normal; the runs that produce neither are the\nones worth being suspicious of.\n","# Run the skill forward tests\n\nThe [forward-test index](..\u002F..\u002F..\u002Fdocs\u002Fskill-forward-tests.md) routes to the\nscenario contracts, measurement dashboard, run archive, and findings worklist.\nThe scenario documents are the contract; this skill is how to execute them\nwithout rediscovering the operational traps each time.\n\nThe index also routes to the deliberately expensive end-to-end journey. Use\nthat journey only for cross-lifecycle changes or release-candidate confidence;\ndo not add it to the ordinary per-skill batch.\n\nThis tests **product-managed skills** — the ones under\n`tools\u002Fspecbind\u002Fassets\u002Fskills\u002F` that get installed into consumer projects. It is\nnot about this repository's own workflow.\n\n## Before anything\n\nRecord the commit under test:\n\n```sh\ngit rev-parse --short HEAD\n```\n\nA run measures one build. If you fix something mid-run, every scenario still in\nflight is measuring the old build, and its result is about that build only. Say\nso rather than folding it into the same table.\n\n## One fixture per scenario\n\n```sh\nsh tools\u002Fspecbind\u002Fscripts\u002Fforward-test-fixture.sh \u002Ftmp\u002Fsb-\u003Cscenario> en\n```\n\nNever share a fixture between scenarios. Several depend on the starting state,\nand a leftover milestone silently changes what is being tested. The builder\nrefuses to touch an existing directory, so delete and rebuild.\n\nThe builder prints an `export PATH=` line for the CLI it ships inside the\nfixture. That line is not optional. The skills invoke `specbind` as a bare\ncommand because a real installed project has it on PATH; without it the run stops\nat \"command not found\" and produces no result about the skill at all.\n\nMost scenarios have a recipe that builds their starting state and proves it:\n\n```sh\nsh tools\u002Fspecbind\u002Fscripts\u002Fforward-test-scenario.sh \u003Cscenario> \u002Ftmp\u002Fsb-\u003Cscenario> en\n```\n\nIt wraps the fixture builder, applies the scenario's precondition, and exits\nnonzero with a message when the precondition did not take. Use it rather than\ncomposing the state by hand; that is where a run silently becomes a different\nrun. Add a recipe when a scenario needs one, and give it a check that fails when\nthe setup is a no-op.\n\nThe end-to-end journey has its own prepare-and-judge harness because it starts\nnear the beginning of the lifecycle and its final state crosses several\nscenario families:\n\n```sh\nsh tools\u002Fspecbind\u002Fscripts\u002Fforward-test-journey.sh prepare hp1 \u002Ftmp\u002Fsb-hp1 en\n# Drive the exact conversation in docs\u002Fskill-forward-tests\u002Fjourney-scenarios.md.\nsh tools\u002Fspecbind\u002Fscripts\u002Fforward-test-journey.sh judge hp1 \u002Ftmp\u002Fsb-hp1\n```\n\nDo not run `judge` after repairing the fixture by hand. A failed judgment is the\nmeasurement. Rebuild at a new target and rerun only when a fresh measurement is\nactually worth its agent cost.\n\nScenarios with no recipe yet need a precondition set up first — an uncommitted edit, a\nbroken steering document. **Do that from the shell**, not from a helper script in\nanother language. On Windows the shell maps `\u002Ftmp` to a real directory, but a\nnative interpreter invoked from it does not, so a script that opens\n`\u002Ftmp\u002Fsb-\u003Cscenario>\u002F...` fails while every shell command beside it succeeds. A\nprecondition that silently did not apply turns a scenario into a different one.\nVerify the precondition with a command before launching the run.\n\nAdd `--instrument-dispatch` as the final builder or recipe argument only when\nthe run is intended to prove fresh-context dispatch. It injects the ignored\n`.forward-test\u002Fagents.log` instruction used to distinguish dispatch from the\nmain-context fallback. Ordinary workflow scenarios leave it off: they have no\ndispatch claim to prove, and an unrelated instrumentation write must not block\nthe product workflow before it starts.\n\n## Driving a run\n\nPrefer a real top-level session through the harness whenever the scenario\ncrosses an approval, dispatches, or measures Skill selection. A subagent cannot\ndo any of those faithfully (ENV-0001, ENV-0003):\n\n```sh\nsh tools\u002Fspecbind\u002Fscripts\u002Fforward-test-drive.sh start \u003Cclaude-code|codex> \u002Ftmp\u002Fsb-\u003Cscenario> \"\u003Cquoted request>\"\nsh tools\u002Fspecbind\u002Fscripts\u002Fforward-test-drive.sh send \u002Ftmp\u002Fsb-\u003Cscenario> \"\u003Cconfirmation>\" --expect '\u003Cboundary check>'\n```\n\nThe harness runs the agent's own CLI in the fixture, removes host-session\nvariables, keeps the maintainer's user-level instructions out, and resumes the\nsame session for each turn. Send only the scenario's quoted request: the\nworking-directory, PATH, and stands-alone statements below are what the harness\nalready makes true. Give every confirmation an `--expect` proving the run\nstopped at that boundary; a failed check sends nothing. Exit 5 is\nenvironment-invalid. Read `\u003Cfixture>.drive\u002Fturns\u002FNN-actions.txt` for which Skill\nthe registry selected, and see the forward-test running guide for the full\ncontract.\n\nOtherwise use a subagent with no prior context.\n\n### Codex driver profile\n\nWhen Codex drives a scenario through a subagent, use this profile:\n\n- `fork_turns: \"none\"`\n- `model: \"gpt-6.1-sol\"`\n- `reasoning_effort: \"medium\"`\n\nThis is the default forward-test driver profile, not part of the product skill\nunder test. Override it only when the run deliberately compares models, and\nrecord the override with the result.\n\nA Codex-spawned subagent is a Codex result; a Claude Code Agent-tool subagent is\na Claude Code result. Record the actual driver because the scenario document\ntracks them as separate columns.\n\nThe prompt gives three things and nothing else:\n\n1. the working directory, given as the **native path** its file tools resolve,\n   with the shell alias named separately if the two differ. On Windows a subagent\n   writing to `\u002Ftmp\u002Fsb-\u003Cscenario>\u002F...` with a file tool can land under the drive\n   root instead of the shell's temp mapping, creating a file the CLI then reports\n   as missing. One run caught that itself and moved the file; the next may not.\n2. the `export PATH=` line, stated as an environment fact\n3. the maintainer's request, phrased the way a maintainer would phrase it\n\nThen ask only what it changed and what it ran.\n\nThe fixture's own `AGENTS.md` or `CLAUDE.md` is applicable project context. Say\nthat explicitly when the driver would otherwise treat a prompted working\ndirectory as less authoritative than the host session. If it never reads the\nfixture instructions and installed skill tree, the product skill was not\nmeasured.\n\nOn Windows, also state the native Git Bash `sh.exe` location when a fixture's\ncanonical command uses `sh`. The printed `export PATH=` line still remains the\nenvironment fact for the CLI. A run that cannot resolve either executable is an\nenvironment failure, not a verdict about the product skill.\n\n**Neutralize this repository's instructions.** A subagent is not a clean room: it\ncarries the host session's project instructions (`AGENTS.md` under Codex or\n`CLAUDE.md` under Claude Code), not the fixture's. This repository's own\ninstructions tell an agent to answer in Japanese and to commit each finished unit\nto `main` — and the checkpoint scenarios exist to measure whether the agent\ncommits. State in the prompt that the fixture is a standalone project and that\ninstructions from any other repository do not apply to it. That is environment\nhygiene, not method.\n\nA run reporting in Japanese against an `en` fixture is the visible symptom. Treat\nit as a signal that the contamination is active, and re-read any checkpoint\nresult taken from the same batch.\n\n**Never name a skill or a command in the prompt.** Whether the agent finds and\nuses the installed skill is the thing under test; telling it teaches the answer.\n\n**Never ask it to justify its classification.** An expectation about what the\nagent told the user cannot be measured from a report you asked for — you get the\nsentence because you requested it. Read that from what the run said on its own.\n\n## Expect a confirmation turn\n\nEvery scenario that crosses a guarded transition has one. Discovery confirms\nscope before its mutation; authoring phases may write a draft first and then\nconfirm its approval. A first pass that stops at the applicable boundary is the\nskill working. Judge the draft and guarded state appropriate to that phase\ninstead of assuming the whole repository must still be unchanged.\n\nAnswer as the maintainer would, then let it finish. Continue the same subagent\nrather than starting a new one, so it keeps the state it gathered. The\nconfirm-then-mutate path is part of what the scenario tests, and it is where the\nguarded CLI operations actually run.\n\n**Confirm the phase, not the project.** Refer explicitly to what the agent just\npresented and where to stop: \"I approve the plan you just presented for Discovery\nonly. Stop after Discovery.\" or \"I approve the Requirements and active Requirement\nID selection you just presented. Stop after Requirements.\" If an authoring phase\ndid not present every value its approval accepts, do not supply or infer the missing\nvalue for it; the scenario has not reached an approvable boundary. A bare \"Proceed\nwith that scope\" can read as permission\nto build the whole thing: one run answered that way went from discovery through\nrequirements, design, review, tasks, and implementation, which left every\ndiscovery expectation unmeasurable because later phases had legitimately changed\nthe same files.\n\n## Judge from the fixture, never from the report\n\nRead the result with commands against the fixture:\n\n```sh\nspecbind milestone status\nspecbind milestone scope\nspecbind spec list\nspecbind spec status \u003Cspec>\ngit status --short\n```\n\nA subagent's summary is a claim about what it did. It is usually accurate and it\nis not evidence. Every expectation in the scenario document is readable from the\nfixture; read it there.\n\n## Ask for a usability debrief after judgment\n\nOnly after the fixture has been judged and the pass or failure recorded, ask the\nsame driver for a qualitative usability debrief. Asking earlier contaminates the\nscenario: it teaches the agent which surfaces the maintainer cares about and can\nturn a naturally discovered ambiguity into a prompted answer.\n\nRecord `git status --short` before the debrief. Then continue the same driver\nwith this read-only follow-up; do not start a fresh agent, because the useful\nevidence is what this agent remembers hesitating over:\n\n```text\nThe scenario result has already been judged. Do not run commands or change any\nfiles. Reflect only on the work you just completed.\n\nWhat made you hesitate, infer an unstated fact, take an unnecessary extra step,\nor risk choosing the wrong action? Consider the CLI, skill, template, protocol,\nadapter, diagnostics, and any other product surface that actually affected the\nwork. Do not invent an issue for every category. If there was no friction, say\n\"none\".\n\nFor each observation, report only:\n- Surface: CLI | Skill | Template | Protocol | Adapter | Other\n- Friction: what was difficult or ambiguous\n- Evidence: the exact command, output, wording, or path you encountered\n- Workaround: what you inferred or did to continue\n- Impact: cosmetic | extra-step | ambiguity | wrong-action-risk\n```\n\nAfter the answer, run `git status --short` again. A changed fixture means the\ndebrief violated its read-only boundary; discard the observation, classify the\nmutation before cleanup, and do not fold it into the scenario result.\n\nThe debrief is not pass\u002Ffail evidence and does not overturn a result already\nmeasured from the fixture. It is an exploratory observation. Record it\nseparately, reproduce it against the owning contract before changing the\nproduct, and look for recurrence across scenarios or agents. One\n`wrong-action-risk` is enough to investigate before the next batch; lower-impact\nobservations normally need either concrete reproduction or a repeated pattern.\n\nTreat `docs\u002Fskill-forward-tests\u002Ffindings.md` as a triaged worklist, not an\nappend-only transcript. Keep reproduced unresolved product findings, compact\nresolved findings to their behavior change and fixing build, and retain\nenvironment limitations only while they affect interpretation. Remove\nduplicates, one-off non-defects, fixture-only workarounds, and `none`\nobservations from the worklist after the run itself records their disposition.\n\n## When something fails\n\nDecide which of these it is before changing anything:\n\n- **The skill is wrong.** Fix the skill.\n- **The decision is wrong.** Fix the decision first, then the skill. This is the\n  common case when a scenario and an accepted decision disagree, and it is why\n  the scenario document says to check the decision before editing a row.\n- **The scenario is wrong.** Fix it, but only after confirming the decisions do\n  not require what it asked for.\n- **The harness is wrong.** Fix the fixture builder or this skill, and note that\n  the scenario never actually ran.\n\nNever adjust an expectation until it passes. A forward test tuned to pass has\nstopped measuring anything.\n\nA scenario that fails once and passes on retry is a finding, not a flake: the\nskill is ambiguous enough for the agent to go either way, and that ambiguity is\nthe defect.\n\nThe same applies across agents. Both are handed the identical skill body —\nrendering only rewrites Front Matter — so a scenario that passes under one and\nfails under the other proves the document admits two readings. That is a skill\ndefect, and \"the other agent behaves differently\" restates it rather than\nexplains it.\n\n## Finishing\n\nCopy `docs\u002Fskill-forward-tests\u002Frun-template.md` to\n`docs\u002Fskill-forward-tests\u002Fruns\u002FYYYY-MM-DD-\u003Cdriver>-\u003Cshort-build>[-N].md`. One\nrecord contains one driver and one tested build. Use `pass`,\n`product_failure`, `scenario_invalid`, `environment_invalid`, or\n`environment_blocked` per scenario, and record the expectation that did not\nhold, the state left behind, and fixture evidence. A retry never replaces its\nfailed attempt.\n\nUpdate `docs\u002Fskill-forward-tests\u002Fresults.md` only with the current projection.\nGive reproduced product findings stable `FT-NNNN` identities in\n`docs\u002Fskill-forward-tests\u002Ffindings.md`; use separate `ENV-NNNN` identities for\nactive environment limitations. Record non-findings and discarded observations\nonly in the run record. Then clean up the fixtures.\n\nRecord any defect the run exposed against the decision or skill it belongs to,\nand update the scenario document when the run showed the procedure itself was\nwrong. Both kinds of finding are normal; the runs that produce neither are the\nones worth being suspicious of.\n",{"name":5,"description":7},"main","2259916910e45633c21f5760d7fe329addc415a7",".agents\u002Fskills\u002Fsb-dev-forward-test",0,"https:\u002F\u002Fgithub.com\u002FHuruikagi\u002Fspecbind","https:\u002F\u002Fgithub.com\u002FHuruikagi","2026-11-04 13:59:47",{"source":8,"skillName":5,"commands":32,"prompt":42,"zipAvailable":43},[33,36,39],{"tool":34,"command":35},"npx","npx skills add Huruikagi\u002Fspecbind --skill sb-dev-forward-test",{"tool":37,"command":38},"bunx","bunx skills add Huruikagi\u002Fspecbind --skill sb-dev-forward-test",{"tool":40,"command":41},"pnpm","pnpm dlx skills add Huruikagi\u002Fspecbind --skill sb-dev-forward-test","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002FHuruikagi\u002Fspecbind\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fsb-dev-forward-test\u002FSKILL.md",true,[45],{"path":46,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":49,"sourceHtmlUrl":50},"agents\u002Fopenai.yaml","REFERENCE","参考文档",121,"https:\u002F\u002Fgithub.com\u002FHuruikagi\u002Fspecbind\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fsb-dev-forward-test\u002Fagents\u002Fopenai.yaml",[52,58,63,70,77,84],{"skillNo":53,"name":54,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":55,"categories":56,"keywords":57,"publishedAt":19},"01M7AH3VQ2VR2X159H99G73TGE","kiro-spec-init","https:\u002F\u002Fgithub.com\u002FHuruikagi\u002Fspecbind\u002Fblob\u002Fmain\u002Ftools\u002Fspecbind\u002Ftests\u002Ffixtures\u002Fmigration\u002Fminimal\u002F.agents\u002Fskills\u002Fkiro-spec-init\u002FSKILL.md",[],[],{"skillNo":59,"name":54,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":60,"categories":61,"keywords":62,"publishedAt":19},"01M7AH3RXXAHA50BR132P0M955","https:\u002F\u002Fgithub.com\u002FHuruikagi\u002Fspecbind\u002Fblob\u002Fmain\u002Ftools\u002Fspecbind\u002Ftests\u002Ffixtures\u002Fmigration\u002Fguided\u002F.claude\u002Fskills\u002Fkiro-spec-init\u002FSKILL.md",[],[],{"skillNo":64,"name":65,"nameZh":6,"description":66,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":67,"categories":68,"keywords":69,"publishedAt":19},"01M7AH3KZBF4YBG4Q8628WD50S","sb-verify-completion","Consequence-free check of an explicit completion or success claim against fresh evidence. Changes nothing; do not use to advance a named Spec whose implementation may be accepted.","https:\u002F\u002Fgithub.com\u002FHuruikagi\u002Fspecbind\u002Fblob\u002Fmain\u002Ftools\u002Fspecbind\u002Fassets\u002Fskills\u002Fsb-verify-completion\u002FSKILL.md",[],[],{"skillNo":71,"name":72,"nameZh":6,"description":73,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":74,"categories":75,"keywords":76,"publishedAt":19},"01M7AH3EFD7KC9RGPZFZE0YNA6","sb-validate-implementation","Validate completed Tasks against active Requirement IDs and record completion on GO under an explicit lifecycle-validation request or ordinary Drive authority. Also owns completion withdrawal under an evidenced Drive replan recovery handoff. For a consequence-free claim check, use sb-verify-completion.","https:\u002F\u002Fgithub.com\u002FHuruikagi\u002Fspecbind\u002Fblob\u002Fmain\u002Ftools\u002Fspecbind\u002Fassets\u002Fskills\u002Fsb-validate-implementation\u002FSKILL.md",[],[],{"skillNo":78,"name":79,"nameZh":6,"description":80,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":81,"categories":82,"keywords":83,"publishedAt":19},"01M7AH38XP0RT7N8G5M74N7FD5","sb-validate-design","Independently judge whether a Spec's design is ready to be built on — requirement coverage, boundary, buildability, self-containment, and architectural fit.","https:\u002F\u002Fgithub.com\u002FHuruikagi\u002Fspecbind\u002Fblob\u002Fmain\u002Ftools\u002Fspecbind\u002Fassets\u002Fskills\u002Fsb-validate-design\u002FSKILL.md",[],[],{"skillNo":85,"name":86,"nameZh":6,"description":87,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":88,"categories":89,"keywords":90,"publishedAt":19},"01M7AH344YTCDM8E8FTSJHC8WF","sb-steering","Maintain durable project guidance — bootstrap it, synchronize it after code changes, or document a long-lived project convention such as testing, API, security, or deployment.","https:\u002F\u002Fgithub.com\u002FHuruikagi\u002Fspecbind\u002Fblob\u002Fmain\u002Ftools\u002Fspecbind\u002Fassets\u002Fskills\u002Fsb-steering\u002FSKILL.md",[],[]]