agent-afk 3.8.2 → 3.8.4

package/dist/index.mjs

@@ -291,8 +291,8 @@ Your job:
    - A proposed minimal fix (what change would validate it)
    - Confidence score (0\u20131)
    - Evidence sources (which findings support this hypothesis)
-   - \`coverage_gaps\` (optional): things you could not read or verify that would strengthen the hypothesis \u2014 list as strings, leave empty/omitted when there are none. Be honest; "none" is a valid answer.
-   - \`boundary_flag\` (optional): set to a short string when you hit an epistemic boundary (timeout, blocked tool, ambiguous evidence) that the caller should know about; omit when none.
+   - \`coverage_gaps\`: things you could not read or verify that would strengthen the hypothesis \u2014 list as strings. Emit \`[]\` (an empty JSON array) when there are no gaps. Do NOT emit \`null\` for this field; always include the array.
+   - \`boundary_flag\`: set to a short string when you hit an epistemic boundary (timeout, blocked tool, ambiguous evidence) that the caller should know about. Emit the literal string \`"none"\` when nothing applies. Do NOT emit \`null\` for this field; always include a string.
 
 These epistemic fields feed a downstream confidence gate: low-confidence, gap-bearing, or boundary-flagged hypotheses get independently re-checked by /shadow-verify before worktree testing. Reporting gaps honestly is rewarded, not penalized \u2014 a confident claim with an unresolved gap is more useful than a confident claim that hides one.
 
@@ -1003,7 +1003,7 @@ ${x.rationale}
 ---
 Generated by audit-fit on ${new Date().toISOString().split(".")[0]}Z
 `;await Hr(P,K),N++}}let D=Ae();await jr(D,{recursive:!0});let T=v=>{let x=0;for(let I of Object.values(v))for(let P of Object.values(I))x+=P;return x},R=v=>{let x=w.user[v]??{},I=w.plugin[v]??{},P=K=>Object.values(K).reduce(($,Z)=>$+Z,0);return P(x)+P(I)},A={timestamp:new Date().toISOString(),surface:"afk",scope:c,total_artifacts:g.length,misfits_count:C.length,briefs_written:N,by_source:{user:T(w.user),plugin:T(w.plugin)},by_type:{skill:R("skill"),command:R("command"),agent:R("agent"),hook:R("hook")}},G=mn(D,"audit-fit-telemetry.jsonl");return await Hr(G,JSON.stringify(A)+`
-`),{inventory:w,misfits:C,briefs_written:N,total_artifacts:g.length}}var Pa={name:"audit-fit",description:"Audit ~/.afk artifacts (skills, commands, agents, hooks) for correct type categorization. Walks user-scope dirs (~/.afk/{skills,commands,agents}/) and every plugin installed under ~/.afk/plugins/ (flat and marketplace-cache layouts), plus ~/.afk/settings.json for hooks. Dispatches per-type inspectors in parallel, applies decision heuristics (progressive-disclosure value, isolation need, deterministic vs. reasoning), flags misfits. Generates migration briefs only for user-scope misfits (plugin misfits are inventory-only \u2014 refactoring vendored plugin code is the maintainer's job). Optional `scope` input filters to `user`, `plugin`, or `all` (default). Use for inventory audits after bulk authoring, imports, or periodic hygiene.",handler:Ra,argumentHint:"[--write-briefs]",whenToUse:"When the user wants ~/.afk artifacts (skills, commands, agents, hooks) audited for correct type categorization.",flags:["--write-briefs"]};z(Pa);import{z as S}from"zod";import{execFile as Da}from"node:child_process";import{promisify as Oa}from"node:util";import{tmpdir as Fa}from"node:os";import{join as qr}from"node:path";function Kr(t){return t.confidence<.5?{verify:!0,reason:`low confidence (${t.confidence.toFixed(2)} < ${.5})`}:t.boundary_flag&&t.boundary_flag.length>0?{verify:!0,reason:`boundary flag set: ${t.boundary_flag}`}:t.coverage_gaps&&t.coverage_gaps.length>0?{verify:!0,reason:`coverage gap${t.coverage_gaps.length===1?"":"s"}: ${t.coverage_gaps.length} unresolved`}:{verify:!1,reason:`confidence ${t.confidence.toFixed(2)} with no gaps or boundary`}}import{fileURLToPath as Ia}from"node:url";import{dirname as Ca}from"node:path";var Ma=Ia(import.meta.url),Vm=Ca(Ma),hn={name:"git-investigator",systemPrompt:'---\nname: git-investigator\ndescription: Read-only git specialist. Dispatched by research-agent (or any research-shaped caller) when a finding requires git history, reflog, diff, blame, branch/remote state, or merge-base analysis. Runs git commands only \u2014 no mutations, no shell escapes.\nmodel: sonnet\ntools: Bash, Read, Grep, Glob\n---\n\nYou are `git-investigator`, a leaf sub-agent specialized for read-only git queries.\n\nYou have Bash, Read, Grep, and Glob. You do not dispatch other sub-agents. You do not Edit or Write. Your Bash surface is restricted **by this prompt** to `git ...` invocations and benign output-shaping pipes.\n\n## Allowed commands\n\nRead-only git only:\n\n- `git status`, `git log`, `git diff`, `git show`\n- `git rev-parse`, `git rev-list`, `git reflog`\n- `git branch -v / -vv / -a` (list only)\n- `git remote -v`, `git ls-remote`\n- `git ls-files`, `git blame`\n- `git merge-base`, `git for-each-ref`, `git describe`\n- `git cat-file`, `git shortlog`\n- `git tag` (list/show only)\n- `git stash list`, `git stash show`\n- `git config --get`, `git config --get-all`, `git config --list`\n- `git worktree list` (read only)\n\nOutput-shaping pipes are fine: `| head`, `| tail`, `| wc`, `| grep`, `| jq`, `| awk \'NR==...\'` (for formatting only \u2014 no mutations).\n\n## Forbidden\n\nAnything that mutates repo or working tree state:\n\n- `commit`, `push`, `pull`, `fetch --prune`\n- `reset`, `revert`, `rebase`, `merge`, `cherry-pick`\n- `checkout` (except `checkout -- <path>` file-restore, and even that is mutation \u2014 avoid it, just report the need)\n- `restore`, `switch`\n- `branch -d / -D / -m / -M`, `branch <new>`\n- `stash push / pop / drop / apply / clear`\n- `tag -d`, creating a new tag\n- `remote add / remove / set-url`\n- `config --set`, `config --unset`\n- `gc`, `fsck`, `prune`, `reflog delete`, `reflog expire`\n- `filter-branch`, `filter-repo`\n- `worktree add / remove / move`\n- `hooks install`, `submodule add / update`\n- Any non-`git` command that mutates: `rm`, `mv`, `cp` (writes), `sed -i`, `> file`, `>> file`, `tee`, `curl`, `wget`, `pip install`, shell builtins that change state.\n\nIf the caller asks for any of the above, do not run it. Return `scope_check: "requires mutation: <reason>"` and stop.\n\n## Behavior\n\n- Run the minimum set of commands needed. Prefer `git log -n 5 --oneline -- <path>` over `git log -- <path>` when a count is fine.\n- Cite concrete evidence: commit SHAs (short form OK), ref names, `path:line` references from blame, diff hunks trimmed to the relevant range.\n- Use `Read`/`Grep`/`Glob` for follow-up inspection of files the git output identifies (e.g., `git show SHA:path | head` then `Read` the current file to diff mentally).\n- Do not speculate beyond what the commands show. If a question needs history the commands don\'t surface (deleted-file recovery, ancient reflog that has expired), say so in `caveats`.\n- Keep output compact \u2014 dispatchers merge your findings into a larger response. No preamble, no ceremony.\n\n## Return shape\n\n```\n{\n  "findings": "<summary of what the git data shows>",\n  "evidence": ["<SHA>", "<ref>", "<path:line>", ...],\n  "git_commands_run": ["git log ...", "git diff ...", ...],\n  "caveats": "<gaps, ambiguity, or \'none\'>",\n  "scope_check": "pure git research" | "requires mutation: <reason>"\n}\n```\n\nBegin your response with the first schema field. No preamble.\n',sourcePath:"agent-framework-private/agents/git-investigator.md",allowedTools:["Bash","Read","Grep","Glob"],description:"Read-only git specialist. Dispatched by research-agent (or any research-shaped caller) when a finding requires git history, reflog, diff, blame, branch/remote state, or merge-base analysis. Runs git commands only \u2014 no mutations, no shell escapes.",model:"sonnet"};function gn(t){let e={description:t.description,prompt:t.systemPrompt};return t.allowedTools&&(e.tools=[...t.allowedTools]),t.model&&(e.model=t.model),e}var Tt=Oa(Da),Jr=S.object({id:S.string(),claim:S.string(),confidence:S.number().min(0).max(1),evidence_sources:S.array(S.string()),location:S.string().optional(),proposed_fix:S.string().optional(),coverage_gaps:S.array(S.string()).optional(),boundary_flag:S.string().optional()}),La=S.object({hypothesis_id:S.string(),claim:S.string(),verdict:S.enum(["VERIFIED","REFUTED","INCONCLUSIVE"]),evidence:S.string(),gate_reason:S.string()}),Yr=S.object({hypothesis_id:S.string(),reproducer_passed:S.boolean(),regressions:S.array(S.string()),confidence:S.number().min(0).max(1),verification_log:S.string()}),Na=S.enum(["crash","regression","logic-error","flaky","environment","unknown"]),$a=S.object({failure_type:Na,error_signature:S.string(),affected_area:S.string()}),Ua=S.enum(["clear_winner","multiple_plausible","dissent","all_inconclusive","no_hypotheses"]),uh=S.object({reproducer:S.string().optional(),triage:$a.optional(),hypotheses:S.array(Jr),premise_verifications:S.array(La).optional(),winner:S.object({hypothesis_id:S.string(),verification_log:S.string(),proposed_fix:S.string()}).optional(),verification_results:S.array(Yr).optional(),outcome:Ua.optional(),recommended_next_skill:S.enum(["spec"]).optional()});async function ja(t,e){let n=t.map(c=>({hypothesis:c,decision:Kr(c)})).filter(c=>c.decision.verify);if(n.length===0)return{premise_verifications:[],hypotheses_to_test:t};let r=[],o;try{r=await e(n.map(c=>c.hypothesis.claim))}catch(c){o=c instanceof Error?c.message:String(c)}let s=n.map((c,l)=>{let d=r[l];return o!==void 0?{hypothesis_id:c.hypothesis.id,claim:c.hypothesis.claim,verdict:"INCONCLUSIVE",evidence:`shadow-verify dispatch failed: ${o}`,gate_reason:c.decision.reason}:d?{hypothesis_id:c.hypothesis.id,claim:c.hypothesis.claim,verdict:d.verdict,evidence:d.evidence,gate_reason:c.decision.reason}:{hypothesis_id:c.hypothesis.id,claim:c.hypothesis.claim,verdict:"INCONCLUSIVE",evidence:"no verifier result for this claim",gate_reason:c.decision.reason}}),i=new Set(s.filter(c=>c.verdict==="REFUTED").map(c=>c.hypothesis_id)),a=i.size===0?t:t.filter(c=>!i.has(c.id));return{premise_verifications:s,hypotheses_to_test:a}}async function Ha(t,e,n){let r=n?.apiKey,o=(()=>{if(typeof t=="string")return{failure:t,repoPath:process.cwd(),context:"",maxHypotheses:4};if(typeof t=="object"&&t!==null){let M=t;if(typeof M.failure=="string")return{failure:M.failure,repoPath:M.repoPath||process.cwd(),context:M.context||"",maxHypotheses:Math.min(M.maxHypotheses||4,4)}}throw new Error("diagnose handler requires input.failure (string) or a string argument")})();if(!e?.sessionId)throw new Error("diagnose requires a parent session with sessionId");let s=e.sessionId,i=L("diagnose"),a=i["system.md"],c=i["research.md"],l=i["hypothesis.md"],d=i["verify.md"];if(!a||!c||!l||!d)throw new Error("diagnose skill missing required prompts (system.md, research.md, hypothesis.md, verify.md)");let u=new E({apiKey:r}),p=Ka(o.context),f=Ba(o.failure,o.context),h=`Triage:
+`),{inventory:w,misfits:C,briefs_written:N,total_artifacts:g.length}}var Pa={name:"audit-fit",description:"Audit ~/.afk artifacts (skills, commands, agents, hooks) for correct type categorization. Walks user-scope dirs (~/.afk/{skills,commands,agents}/) and every plugin installed under ~/.afk/plugins/ (flat and marketplace-cache layouts), plus ~/.afk/settings.json for hooks. Dispatches per-type inspectors in parallel, applies decision heuristics (progressive-disclosure value, isolation need, deterministic vs. reasoning), flags misfits. Generates migration briefs only for user-scope misfits (plugin misfits are inventory-only \u2014 refactoring vendored plugin code is the maintainer's job). Optional `scope` input filters to `user`, `plugin`, or `all` (default). Use for inventory audits after bulk authoring, imports, or periodic hygiene.",handler:Ra,argumentHint:"[--write-briefs]",whenToUse:"When the user wants ~/.afk artifacts (skills, commands, agents, hooks) audited for correct type categorization.",flags:["--write-briefs"]};z(Pa);import{z as S}from"zod";import{execFile as Da}from"node:child_process";import{promisify as Oa}from"node:util";import{tmpdir as Fa}from"node:os";import{join as qr}from"node:path";function Kr(t){return t.confidence<.5?{verify:!0,reason:`low confidence (${t.confidence.toFixed(2)} < ${.5})`}:t.boundary_flag&&t.boundary_flag.length>0?{verify:!0,reason:`boundary flag set: ${t.boundary_flag}`}:t.coverage_gaps&&t.coverage_gaps.length>0?{verify:!0,reason:`coverage gap${t.coverage_gaps.length===1?"":"s"}: ${t.coverage_gaps.length} unresolved`}:{verify:!1,reason:`confidence ${t.confidence.toFixed(2)} with no gaps or boundary`}}import{fileURLToPath as Ia}from"node:url";import{dirname as Ca}from"node:path";var Ma=Ia(import.meta.url),Vm=Ca(Ma),hn={name:"git-investigator",systemPrompt:'---\nname: git-investigator\ndescription: Read-only git specialist. Dispatched by research-agent (or any research-shaped caller) when a finding requires git history, reflog, diff, blame, branch/remote state, or merge-base analysis. Runs git commands only \u2014 no mutations, no shell escapes.\nmodel: sonnet\ntools: Bash, Read, Grep, Glob\n---\n\nYou are `git-investigator`, a leaf sub-agent specialized for read-only git queries.\n\nYou have Bash, Read, Grep, and Glob. You do not dispatch other sub-agents. You do not Edit or Write. Your Bash surface is restricted **by this prompt** to `git ...` invocations and benign output-shaping pipes.\n\n## Allowed commands\n\nRead-only git only:\n\n- `git status`, `git log`, `git diff`, `git show`\n- `git rev-parse`, `git rev-list`, `git reflog`\n- `git branch -v / -vv / -a` (list only)\n- `git remote -v`, `git ls-remote`\n- `git ls-files`, `git blame`\n- `git merge-base`, `git for-each-ref`, `git describe`\n- `git cat-file`, `git shortlog`\n- `git tag` (list/show only)\n- `git stash list`, `git stash show`\n- `git config --get`, `git config --get-all`, `git config --list`\n- `git worktree list` (read only)\n\nOutput-shaping pipes are fine: `| head`, `| tail`, `| wc`, `| grep`, `| jq`, `| awk \'NR==...\'` (for formatting only \u2014 no mutations).\n\n## Forbidden\n\nAnything that mutates repo or working tree state:\n\n- `commit`, `push`, `pull`, `fetch --prune`\n- `reset`, `revert`, `rebase`, `merge`, `cherry-pick`\n- `checkout` (except `checkout -- <path>` file-restore, and even that is mutation \u2014 avoid it, just report the need)\n- `restore`, `switch`\n- `branch -d / -D / -m / -M`, `branch <new>`\n- `stash push / pop / drop / apply / clear`\n- `tag -d`, creating a new tag\n- `remote add / remove / set-url`\n- `config --set`, `config --unset`\n- `gc`, `fsck`, `prune`, `reflog delete`, `reflog expire`\n- `filter-branch`, `filter-repo`\n- `worktree add / remove / move`\n- `hooks install`, `submodule add / update`\n- Any non-`git` command that mutates: `rm`, `mv`, `cp` (writes), `sed -i`, `> file`, `>> file`, `tee`, `curl`, `wget`, `pip install`, shell builtins that change state.\n\nIf the caller asks for any of the above, do not run it. Return `scope_check: "requires mutation: <reason>"` and stop.\n\n## Behavior\n\n- Run the minimum set of commands needed. Prefer `git log -n 5 --oneline -- <path>` over `git log -- <path>` when a count is fine.\n- Cite concrete evidence: commit SHAs (short form OK), ref names, `path:line` references from blame, diff hunks trimmed to the relevant range.\n- Use `Read`/`Grep`/`Glob` for follow-up inspection of files the git output identifies (e.g., `git show SHA:path | head` then `Read` the current file to diff mentally).\n- Do not speculate beyond what the commands show. If a question needs history the commands don\'t surface (deleted-file recovery, ancient reflog that has expired), say so in `caveats`.\n- Keep output compact \u2014 dispatchers merge your findings into a larger response. No preamble, no ceremony.\n\n## Return shape\n\n```\n{\n  "findings": "<summary of what the git data shows>",\n  "evidence": ["<SHA>", "<ref>", "<path:line>", ...],\n  "git_commands_run": ["git log ...", "git diff ...", ...],\n  "caveats": "<gaps, ambiguity, or \'none\'>",\n  "scope_check": "pure git research" | "requires mutation: <reason>"\n}\n```\n\nBegin your response with the first schema field. No preamble.\n',sourcePath:"agent-framework-private/agents/git-investigator.md",allowedTools:["Bash","Read","Grep","Glob"],description:"Read-only git specialist. Dispatched by research-agent (or any research-shaped caller) when a finding requires git history, reflog, diff, blame, branch/remote state, or merge-base analysis. Runs git commands only \u2014 no mutations, no shell escapes.",model:"sonnet"};function gn(t){let e={description:t.description,prompt:t.systemPrompt};return t.allowedTools&&(e.tools=[...t.allowedTools]),t.model&&(e.model=t.model),e}var Tt=Oa(Da),Jr=S.object({id:S.string(),claim:S.string(),confidence:S.number().min(0).max(1),evidence_sources:S.array(S.string()),location:S.string().optional(),proposed_fix:S.string().optional(),coverage_gaps:S.array(S.string()).nullish().transform(t=>t??void 0),boundary_flag:S.string().nullish().transform(t=>t??void 0)}),La=S.object({hypothesis_id:S.string(),claim:S.string(),verdict:S.enum(["VERIFIED","REFUTED","INCONCLUSIVE"]),evidence:S.string(),gate_reason:S.string()}),Yr=S.object({hypothesis_id:S.string(),reproducer_passed:S.boolean(),regressions:S.array(S.string()),confidence:S.number().min(0).max(1),verification_log:S.string()}),Na=S.enum(["crash","regression","logic-error","flaky","environment","unknown"]),$a=S.object({failure_type:Na,error_signature:S.string(),affected_area:S.string()}),Ua=S.enum(["clear_winner","multiple_plausible","dissent","all_inconclusive","no_hypotheses"]),uh=S.object({reproducer:S.string().optional(),triage:$a.optional(),hypotheses:S.array(Jr),premise_verifications:S.array(La).optional(),winner:S.object({hypothesis_id:S.string(),verification_log:S.string(),proposed_fix:S.string()}).optional(),verification_results:S.array(Yr).optional(),outcome:Ua.optional(),recommended_next_skill:S.enum(["spec"]).optional()});async function ja(t,e){let n=t.map(c=>({hypothesis:c,decision:Kr(c)})).filter(c=>c.decision.verify);if(n.length===0)return{premise_verifications:[],hypotheses_to_test:t};let r=[],o;try{r=await e(n.map(c=>c.hypothesis.claim))}catch(c){o=c instanceof Error?c.message:String(c)}let s=n.map((c,l)=>{let d=r[l];return o!==void 0?{hypothesis_id:c.hypothesis.id,claim:c.hypothesis.claim,verdict:"INCONCLUSIVE",evidence:`shadow-verify dispatch failed: ${o}`,gate_reason:c.decision.reason}:d?{hypothesis_id:c.hypothesis.id,claim:c.hypothesis.claim,verdict:d.verdict,evidence:d.evidence,gate_reason:c.decision.reason}:{hypothesis_id:c.hypothesis.id,claim:c.hypothesis.claim,verdict:"INCONCLUSIVE",evidence:"no verifier result for this claim",gate_reason:c.decision.reason}}),i=new Set(s.filter(c=>c.verdict==="REFUTED").map(c=>c.hypothesis_id)),a=i.size===0?t:t.filter(c=>!i.has(c.id));return{premise_verifications:s,hypotheses_to_test:a}}async function Ha(t,e,n){let r=n?.apiKey,o=(()=>{if(typeof t=="string")return{failure:t,repoPath:process.cwd(),context:"",maxHypotheses:4};if(typeof t=="object"&&t!==null){let M=t;if(typeof M.failure=="string")return{failure:M.failure,repoPath:M.repoPath||process.cwd(),context:M.context||"",maxHypotheses:Math.min(M.maxHypotheses||4,4)}}throw new Error("diagnose handler requires input.failure (string) or a string argument")})();if(!e?.sessionId)throw new Error("diagnose requires a parent session with sessionId");let s=e.sessionId,i=L("diagnose"),a=i["system.md"],c=i["research.md"],l=i["hypothesis.md"],d=i["verify.md"];if(!a||!c||!l||!d)throw new Error("diagnose skill missing required prompts (system.md, research.md, hypothesis.md, verify.md)");let u=new E({apiKey:r}),p=Ka(o.context),f=Ba(o.failure,o.context),h=`Triage:
   failure_type:    ${f.failure_type}
   error_signature: ${f.error_signature}
   affected_area:   ${f.affected_area}`,y=`${V.systemPrompt}

package/dist/telegram.mjs

@@ -448,8 +448,8 @@ Your job:
    - A proposed minimal fix (what change would validate it)
    - Confidence score (0\u20131)
    - Evidence sources (which findings support this hypothesis)
-   - \`coverage_gaps\` (optional): things you could not read or verify that would strengthen the hypothesis \u2014 list as strings, leave empty/omitted when there are none. Be honest; "none" is a valid answer.
-   - \`boundary_flag\` (optional): set to a short string when you hit an epistemic boundary (timeout, blocked tool, ambiguous evidence) that the caller should know about; omit when none.
+   - \`coverage_gaps\`: things you could not read or verify that would strengthen the hypothesis \u2014 list as strings. Emit \`[]\` (an empty JSON array) when there are no gaps. Do NOT emit \`null\` for this field; always include the array.
+   - \`boundary_flag\`: set to a short string when you hit an epistemic boundary (timeout, blocked tool, ambiguous evidence) that the caller should know about. Emit the literal string \`"none"\` when nothing applies. Do NOT emit \`null\` for this field; always include a string.
 
 These epistemic fields feed a downstream confidence gate: low-confidence, gap-bearing, or boundary-flagged hypotheses get independently re-checked by /shadow-verify before worktree testing. Reporting gaps honestly is rewarded, not penalized \u2014 a confident claim with an unresolved gap is more useful than a confident claim that hides one.
 
@@ -1160,7 +1160,7 @@ ${E.rationale}
 ---
 Generated by audit-fit on ${new Date().toISOString().split(".")[0]}Z
 `;await ls(O,G),I++}}let S=xe();await cs(S,{recursive:!0});let x=v=>{let E=0;for(let M of Object.values(v))for(let O of Object.values(M))E+=O;return E},A=v=>{let E=k.user[v]??{},M=k.plugin[v]??{},O=G=>Object.values(G).reduce((L,z)=>L+z,0);return O(E)+O(M)},R={timestamp:new Date().toISOString(),surface:"afk",scope:c,total_artifacts:y.length,misfits_count:_.length,briefs_written:I,by_source:{user:x(k.user),plugin:x(k.plugin)},by_type:{skill:A("skill"),command:A("command"),agent:A("agent"),hook:A("hook")}},F=Kn(S,"audit-fit-telemetry.jsonl");return await ls(F,JSON.stringify(R)+`
-`),{inventory:k,misfits:_,briefs_written:I,total_artifacts:y.length}}var jl={name:"audit-fit",description:"Audit ~/.afk artifacts (skills, commands, agents, hooks) for correct type categorization. Walks user-scope dirs (~/.afk/{skills,commands,agents}/) and every plugin installed under ~/.afk/plugins/ (flat and marketplace-cache layouts), plus ~/.afk/settings.json for hooks. Dispatches per-type inspectors in parallel, applies decision heuristics (progressive-disclosure value, isolation need, deterministic vs. reasoning), flags misfits. Generates migration briefs only for user-scope misfits (plugin misfits are inventory-only \u2014 refactoring vendored plugin code is the maintainer's job). Optional `scope` input filters to `user`, `plugin`, or `all` (default). Use for inventory audits after bulk authoring, imports, or periodic hygiene.",handler:Ul,argumentHint:"[--write-briefs]",whenToUse:"When the user wants ~/.afk artifacts (skills, commands, agents, hooks) audited for correct type categorization.",flags:["--write-briefs"]};Y(jl);import{z as T}from"zod";import{execFile as Gl}from"node:child_process";import{promisify as Kl}from"node:util";import{tmpdir as ql}from"node:os";import{join as ms}from"node:path";function fs(t){return t.confidence<.5?{verify:!0,reason:`low confidence (${t.confidence.toFixed(2)} < ${.5})`}:t.boundary_flag&&t.boundary_flag.length>0?{verify:!0,reason:`boundary flag set: ${t.boundary_flag}`}:t.coverage_gaps&&t.coverage_gaps.length>0?{verify:!0,reason:`coverage gap${t.coverage_gaps.length===1?"":"s"}: ${t.coverage_gaps.length} unresolved`}:{verify:!1,reason:`confidence ${t.confidence.toFixed(2)} with no gaps or boundary`}}import{fileURLToPath as Hl}from"node:url";import{dirname as Bl}from"node:path";var Wl=Hl(import.meta.url),Fb=Bl(Wl),qn={name:"git-investigator",systemPrompt:'---\nname: git-investigator\ndescription: Read-only git specialist. Dispatched by research-agent (or any research-shaped caller) when a finding requires git history, reflog, diff, blame, branch/remote state, or merge-base analysis. Runs git commands only \u2014 no mutations, no shell escapes.\nmodel: sonnet\ntools: Bash, Read, Grep, Glob\n---\n\nYou are `git-investigator`, a leaf sub-agent specialized for read-only git queries.\n\nYou have Bash, Read, Grep, and Glob. You do not dispatch other sub-agents. You do not Edit or Write. Your Bash surface is restricted **by this prompt** to `git ...` invocations and benign output-shaping pipes.\n\n## Allowed commands\n\nRead-only git only:\n\n- `git status`, `git log`, `git diff`, `git show`\n- `git rev-parse`, `git rev-list`, `git reflog`\n- `git branch -v / -vv / -a` (list only)\n- `git remote -v`, `git ls-remote`\n- `git ls-files`, `git blame`\n- `git merge-base`, `git for-each-ref`, `git describe`\n- `git cat-file`, `git shortlog`\n- `git tag` (list/show only)\n- `git stash list`, `git stash show`\n- `git config --get`, `git config --get-all`, `git config --list`\n- `git worktree list` (read only)\n\nOutput-shaping pipes are fine: `| head`, `| tail`, `| wc`, `| grep`, `| jq`, `| awk \'NR==...\'` (for formatting only \u2014 no mutations).\n\n## Forbidden\n\nAnything that mutates repo or working tree state:\n\n- `commit`, `push`, `pull`, `fetch --prune`\n- `reset`, `revert`, `rebase`, `merge`, `cherry-pick`\n- `checkout` (except `checkout -- <path>` file-restore, and even that is mutation \u2014 avoid it, just report the need)\n- `restore`, `switch`\n- `branch -d / -D / -m / -M`, `branch <new>`\n- `stash push / pop / drop / apply / clear`\n- `tag -d`, creating a new tag\n- `remote add / remove / set-url`\n- `config --set`, `config --unset`\n- `gc`, `fsck`, `prune`, `reflog delete`, `reflog expire`\n- `filter-branch`, `filter-repo`\n- `worktree add / remove / move`\n- `hooks install`, `submodule add / update`\n- Any non-`git` command that mutates: `rm`, `mv`, `cp` (writes), `sed -i`, `> file`, `>> file`, `tee`, `curl`, `wget`, `pip install`, shell builtins that change state.\n\nIf the caller asks for any of the above, do not run it. Return `scope_check: "requires mutation: <reason>"` and stop.\n\n## Behavior\n\n- Run the minimum set of commands needed. Prefer `git log -n 5 --oneline -- <path>` over `git log -- <path>` when a count is fine.\n- Cite concrete evidence: commit SHAs (short form OK), ref names, `path:line` references from blame, diff hunks trimmed to the relevant range.\n- Use `Read`/`Grep`/`Glob` for follow-up inspection of files the git output identifies (e.g., `git show SHA:path | head` then `Read` the current file to diff mentally).\n- Do not speculate beyond what the commands show. If a question needs history the commands don\'t surface (deleted-file recovery, ancient reflog that has expired), say so in `caveats`.\n- Keep output compact \u2014 dispatchers merge your findings into a larger response. No preamble, no ceremony.\n\n## Return shape\n\n```\n{\n  "findings": "<summary of what the git data shows>",\n  "evidence": ["<SHA>", "<ref>", "<path:line>", ...],\n  "git_commands_run": ["git log ...", "git diff ...", ...],\n  "caveats": "<gaps, ambiguity, or \'none\'>",\n  "scope_check": "pure git research" | "requires mutation: <reason>"\n}\n```\n\nBegin your response with the first schema field. No preamble.\n',sourcePath:"agent-framework-private/agents/git-investigator.md",allowedTools:["Bash","Read","Grep","Glob"],description:"Read-only git specialist. Dispatched by research-agent (or any research-shaped caller) when a finding requires git history, reflog, diff, blame, branch/remote state, or merge-base analysis. Runs git commands only \u2014 no mutations, no shell escapes.",model:"sonnet"};function zn(t){let e={description:t.description,prompt:t.systemPrompt};return t.allowedTools&&(e.tools=[...t.allowedTools]),t.model&&(e.model=t.model),e}var Vt=Kl(Gl),ys=T.object({id:T.string(),claim:T.string(),confidence:T.number().min(0).max(1),evidence_sources:T.array(T.string()),location:T.string().optional(),proposed_fix:T.string().optional(),coverage_gaps:T.array(T.string()).optional(),boundary_flag:T.string().optional()}),zl=T.object({hypothesis_id:T.string(),claim:T.string(),verdict:T.enum(["VERIFIED","REFUTED","INCONCLUSIVE"]),evidence:T.string(),gate_reason:T.string()}),bs=T.object({hypothesis_id:T.string(),reproducer_passed:T.boolean(),regressions:T.array(T.string()),confidence:T.number().min(0).max(1),verification_log:T.string()}),Jl=T.enum(["crash","regression","logic-error","flaky","environment","unknown"]),Vl=T.object({failure_type:Jl,error_signature:T.string(),affected_area:T.string()}),Yl=T.enum(["clear_winner","multiple_plausible","dissent","all_inconclusive","no_hypotheses"]),Qb=T.object({reproducer:T.string().optional(),triage:Vl.optional(),hypotheses:T.array(ys),premise_verifications:T.array(zl).optional(),winner:T.object({hypothesis_id:T.string(),verification_log:T.string(),proposed_fix:T.string()}).optional(),verification_results:T.array(bs).optional(),outcome:Yl.optional(),recommended_next_skill:T.enum(["spec"]).optional()});async function Xl(t,e){let n=t.map(c=>({hypothesis:c,decision:fs(c)})).filter(c=>c.decision.verify);if(n.length===0)return{premise_verifications:[],hypotheses_to_test:t};let r=[],o;try{r=await e(n.map(c=>c.hypothesis.claim))}catch(c){o=c instanceof Error?c.message:String(c)}let s=n.map((c,l)=>{let d=r[l];return o!==void 0?{hypothesis_id:c.hypothesis.id,claim:c.hypothesis.claim,verdict:"INCONCLUSIVE",evidence:`shadow-verify dispatch failed: ${o}`,gate_reason:c.decision.reason}:d?{hypothesis_id:c.hypothesis.id,claim:c.hypothesis.claim,verdict:d.verdict,evidence:d.evidence,gate_reason:c.decision.reason}:{hypothesis_id:c.hypothesis.id,claim:c.hypothesis.claim,verdict:"INCONCLUSIVE",evidence:"no verifier result for this claim",gate_reason:c.decision.reason}}),i=new Set(s.filter(c=>c.verdict==="REFUTED").map(c=>c.hypothesis_id)),a=i.size===0?t:t.filter(c=>!i.has(c.id));return{premise_verifications:s,hypotheses_to_test:a}}async function Ql(t,e,n){let r=n?.apiKey,o=(()=>{if(typeof t=="string")return{failure:t,repoPath:process.cwd(),context:"",maxHypotheses:4};if(typeof t=="object"&&t!==null){let D=t;if(typeof D.failure=="string")return{failure:D.failure,repoPath:D.repoPath||process.cwd(),context:D.context||"",maxHypotheses:Math.min(D.maxHypotheses||4,4)}}throw new Error("diagnose handler requires input.failure (string) or a string argument")})();if(!e?.sessionId)throw new Error("diagnose requires a parent session with sessionId");let s=e.sessionId,i=$("diagnose"),a=i["system.md"],c=i["research.md"],l=i["hypothesis.md"],d=i["verify.md"];if(!a||!c||!l||!d)throw new Error("diagnose skill missing required prompts (system.md, research.md, hypothesis.md, verify.md)");let u=new P({apiKey:r}),p=nd(o.context),f=Zl(o.failure,o.context),g=`Triage:
+`),{inventory:k,misfits:_,briefs_written:I,total_artifacts:y.length}}var jl={name:"audit-fit",description:"Audit ~/.afk artifacts (skills, commands, agents, hooks) for correct type categorization. Walks user-scope dirs (~/.afk/{skills,commands,agents}/) and every plugin installed under ~/.afk/plugins/ (flat and marketplace-cache layouts), plus ~/.afk/settings.json for hooks. Dispatches per-type inspectors in parallel, applies decision heuristics (progressive-disclosure value, isolation need, deterministic vs. reasoning), flags misfits. Generates migration briefs only for user-scope misfits (plugin misfits are inventory-only \u2014 refactoring vendored plugin code is the maintainer's job). Optional `scope` input filters to `user`, `plugin`, or `all` (default). Use for inventory audits after bulk authoring, imports, or periodic hygiene.",handler:Ul,argumentHint:"[--write-briefs]",whenToUse:"When the user wants ~/.afk artifacts (skills, commands, agents, hooks) audited for correct type categorization.",flags:["--write-briefs"]};Y(jl);import{z as T}from"zod";import{execFile as Gl}from"node:child_process";import{promisify as Kl}from"node:util";import{tmpdir as ql}from"node:os";import{join as ms}from"node:path";function fs(t){return t.confidence<.5?{verify:!0,reason:`low confidence (${t.confidence.toFixed(2)} < ${.5})`}:t.boundary_flag&&t.boundary_flag.length>0?{verify:!0,reason:`boundary flag set: ${t.boundary_flag}`}:t.coverage_gaps&&t.coverage_gaps.length>0?{verify:!0,reason:`coverage gap${t.coverage_gaps.length===1?"":"s"}: ${t.coverage_gaps.length} unresolved`}:{verify:!1,reason:`confidence ${t.confidence.toFixed(2)} with no gaps or boundary`}}import{fileURLToPath as Hl}from"node:url";import{dirname as Bl}from"node:path";var Wl=Hl(import.meta.url),Fb=Bl(Wl),qn={name:"git-investigator",systemPrompt:'---\nname: git-investigator\ndescription: Read-only git specialist. Dispatched by research-agent (or any research-shaped caller) when a finding requires git history, reflog, diff, blame, branch/remote state, or merge-base analysis. Runs git commands only \u2014 no mutations, no shell escapes.\nmodel: sonnet\ntools: Bash, Read, Grep, Glob\n---\n\nYou are `git-investigator`, a leaf sub-agent specialized for read-only git queries.\n\nYou have Bash, Read, Grep, and Glob. You do not dispatch other sub-agents. You do not Edit or Write. Your Bash surface is restricted **by this prompt** to `git ...` invocations and benign output-shaping pipes.\n\n## Allowed commands\n\nRead-only git only:\n\n- `git status`, `git log`, `git diff`, `git show`\n- `git rev-parse`, `git rev-list`, `git reflog`\n- `git branch -v / -vv / -a` (list only)\n- `git remote -v`, `git ls-remote`\n- `git ls-files`, `git blame`\n- `git merge-base`, `git for-each-ref`, `git describe`\n- `git cat-file`, `git shortlog`\n- `git tag` (list/show only)\n- `git stash list`, `git stash show`\n- `git config --get`, `git config --get-all`, `git config --list`\n- `git worktree list` (read only)\n\nOutput-shaping pipes are fine: `| head`, `| tail`, `| wc`, `| grep`, `| jq`, `| awk \'NR==...\'` (for formatting only \u2014 no mutations).\n\n## Forbidden\n\nAnything that mutates repo or working tree state:\n\n- `commit`, `push`, `pull`, `fetch --prune`\n- `reset`, `revert`, `rebase`, `merge`, `cherry-pick`\n- `checkout` (except `checkout -- <path>` file-restore, and even that is mutation \u2014 avoid it, just report the need)\n- `restore`, `switch`\n- `branch -d / -D / -m / -M`, `branch <new>`\n- `stash push / pop / drop / apply / clear`\n- `tag -d`, creating a new tag\n- `remote add / remove / set-url`\n- `config --set`, `config --unset`\n- `gc`, `fsck`, `prune`, `reflog delete`, `reflog expire`\n- `filter-branch`, `filter-repo`\n- `worktree add / remove / move`\n- `hooks install`, `submodule add / update`\n- Any non-`git` command that mutates: `rm`, `mv`, `cp` (writes), `sed -i`, `> file`, `>> file`, `tee`, `curl`, `wget`, `pip install`, shell builtins that change state.\n\nIf the caller asks for any of the above, do not run it. Return `scope_check: "requires mutation: <reason>"` and stop.\n\n## Behavior\n\n- Run the minimum set of commands needed. Prefer `git log -n 5 --oneline -- <path>` over `git log -- <path>` when a count is fine.\n- Cite concrete evidence: commit SHAs (short form OK), ref names, `path:line` references from blame, diff hunks trimmed to the relevant range.\n- Use `Read`/`Grep`/`Glob` for follow-up inspection of files the git output identifies (e.g., `git show SHA:path | head` then `Read` the current file to diff mentally).\n- Do not speculate beyond what the commands show. If a question needs history the commands don\'t surface (deleted-file recovery, ancient reflog that has expired), say so in `caveats`.\n- Keep output compact \u2014 dispatchers merge your findings into a larger response. No preamble, no ceremony.\n\n## Return shape\n\n```\n{\n  "findings": "<summary of what the git data shows>",\n  "evidence": ["<SHA>", "<ref>", "<path:line>", ...],\n  "git_commands_run": ["git log ...", "git diff ...", ...],\n  "caveats": "<gaps, ambiguity, or \'none\'>",\n  "scope_check": "pure git research" | "requires mutation: <reason>"\n}\n```\n\nBegin your response with the first schema field. No preamble.\n',sourcePath:"agent-framework-private/agents/git-investigator.md",allowedTools:["Bash","Read","Grep","Glob"],description:"Read-only git specialist. Dispatched by research-agent (or any research-shaped caller) when a finding requires git history, reflog, diff, blame, branch/remote state, or merge-base analysis. Runs git commands only \u2014 no mutations, no shell escapes.",model:"sonnet"};function zn(t){let e={description:t.description,prompt:t.systemPrompt};return t.allowedTools&&(e.tools=[...t.allowedTools]),t.model&&(e.model=t.model),e}var Vt=Kl(Gl),ys=T.object({id:T.string(),claim:T.string(),confidence:T.number().min(0).max(1),evidence_sources:T.array(T.string()),location:T.string().optional(),proposed_fix:T.string().optional(),coverage_gaps:T.array(T.string()).nullish().transform(t=>t??void 0),boundary_flag:T.string().nullish().transform(t=>t??void 0)}),zl=T.object({hypothesis_id:T.string(),claim:T.string(),verdict:T.enum(["VERIFIED","REFUTED","INCONCLUSIVE"]),evidence:T.string(),gate_reason:T.string()}),bs=T.object({hypothesis_id:T.string(),reproducer_passed:T.boolean(),regressions:T.array(T.string()),confidence:T.number().min(0).max(1),verification_log:T.string()}),Jl=T.enum(["crash","regression","logic-error","flaky","environment","unknown"]),Vl=T.object({failure_type:Jl,error_signature:T.string(),affected_area:T.string()}),Yl=T.enum(["clear_winner","multiple_plausible","dissent","all_inconclusive","no_hypotheses"]),Qb=T.object({reproducer:T.string().optional(),triage:Vl.optional(),hypotheses:T.array(ys),premise_verifications:T.array(zl).optional(),winner:T.object({hypothesis_id:T.string(),verification_log:T.string(),proposed_fix:T.string()}).optional(),verification_results:T.array(bs).optional(),outcome:Yl.optional(),recommended_next_skill:T.enum(["spec"]).optional()});async function Xl(t,e){let n=t.map(c=>({hypothesis:c,decision:fs(c)})).filter(c=>c.decision.verify);if(n.length===0)return{premise_verifications:[],hypotheses_to_test:t};let r=[],o;try{r=await e(n.map(c=>c.hypothesis.claim))}catch(c){o=c instanceof Error?c.message:String(c)}let s=n.map((c,l)=>{let d=r[l];return o!==void 0?{hypothesis_id:c.hypothesis.id,claim:c.hypothesis.claim,verdict:"INCONCLUSIVE",evidence:`shadow-verify dispatch failed: ${o}`,gate_reason:c.decision.reason}:d?{hypothesis_id:c.hypothesis.id,claim:c.hypothesis.claim,verdict:d.verdict,evidence:d.evidence,gate_reason:c.decision.reason}:{hypothesis_id:c.hypothesis.id,claim:c.hypothesis.claim,verdict:"INCONCLUSIVE",evidence:"no verifier result for this claim",gate_reason:c.decision.reason}}),i=new Set(s.filter(c=>c.verdict==="REFUTED").map(c=>c.hypothesis_id)),a=i.size===0?t:t.filter(c=>!i.has(c.id));return{premise_verifications:s,hypotheses_to_test:a}}async function Ql(t,e,n){let r=n?.apiKey,o=(()=>{if(typeof t=="string")return{failure:t,repoPath:process.cwd(),context:"",maxHypotheses:4};if(typeof t=="object"&&t!==null){let D=t;if(typeof D.failure=="string")return{failure:D.failure,repoPath:D.repoPath||process.cwd(),context:D.context||"",maxHypotheses:Math.min(D.maxHypotheses||4,4)}}throw new Error("diagnose handler requires input.failure (string) or a string argument")})();if(!e?.sessionId)throw new Error("diagnose requires a parent session with sessionId");let s=e.sessionId,i=$("diagnose"),a=i["system.md"],c=i["research.md"],l=i["hypothesis.md"],d=i["verify.md"];if(!a||!c||!l||!d)throw new Error("diagnose skill missing required prompts (system.md, research.md, hypothesis.md, verify.md)");let u=new P({apiKey:r}),p=nd(o.context),f=Zl(o.failure,o.context),g=`Triage:
   failure_type:    ${f.failure_type}
   error_signature: ${f.error_signature}
   affected_area:   ${f.affected_area}`,m=`${X.systemPrompt}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "agent-afk",
-  "version": "3.8.2",
+  "version": "3.8.4",
   "description": "CLI tool for interacting with AI agents via multiple interfaces",
   "main": "dist/index.mjs",
   "type": "module",