Skip to content

docs: repository-wide accuracy corrections (doc-vs-code audit, v1.35.0) - #164

Merged
jkyberneees merged 1 commit into
mainfrom
fix/security-docs-accuracy
Aug 31, 2026
Merged

docs: repository-wide accuracy corrections (doc-vs-code audit, v1.35.0)#164
jkyberneees merged 1 commit into
mainfrom
fix/security-docs-accuracy

Conversation

@jkyberneees

Copy link
Copy Markdown
Contributor

Summary

Full doc-vs-code accuracy audit: all 24 docs/ files + README.md + AGENTS.md, executed as 9 sub-agent waves (read-only code-reviewer profile, max 2 concurrent) plus a dedicated docs/SECURITY.md deep audit. Every concrete claim — identifiers, flags, env vars, defaults, numeric caps, test names, handler signatures — was checked against source. Docs-only; no code changes.

Notable corrections

File Fix
SECURITY.md / PLANNING.md stale TestIsPlanMessage citations (test does not exist; mechanism lives in refreshPlanMessage); ToolCallTaints is not "shared with skills"; undocumented untoggleable mcp_schema guard scope; plan-files bounds row (actual: maxTelegramPlanChars = 3800); MCP server-mode non_interactive default is read_only, not deny; utf16Len claim removed (counting is byte-based); <system> token requires an override verb
CONFIG.md / CLI.md phantom ODEK_TOOL_PROGRESS env removed; --model default is deepseek-v4-flash; repl rejects unknown flags (not silently ignored); serve has no --no-compaction; "every knob has an env counterpart" softened
MCP.md exposed-tools table +4 (skill_load/skill_list/artifact_read/list_subagent_profiles), plan conditional; approval keys also hash schema+description (sentence was garbled); per-tool approval applies to every server; discovery errors fail fast (not "skip and continue")
TELEGRAM.md /plan_resume takes no args; all 4 handler signatures carry trailing userID int64
SANDBOXING.md / README / AGENTS.md sandbox is default-on for run/repl/serve (--no-sandbox/ODEK_NO_SANDBOX opt-out, ODEK_REQUIRE_SANDBOX fatal); stale render.Start() symbol; per-tool-approval bullet; event-type lists completed (subagent_spawned/subagent_completed)
CHEATSHEET / TOOL_SELECTION memory tool has 16 actions (was 6); phantom skill_save/skill_patch/skill_delete agent tools removed; list_subagent_profiles/artifact_read added
SUBAGENTS / EXTENSIONS / MAINTENANCE / PROVIDERS / DEVELOPMENT / WEBUI / etc. per-line Unmarshal, max_concurrency fallback documented, artifacts_max_age_hours in config reference, 11 danger classes, docs/ index completed (+14 entries), version example freshened
DOCKER_COMPOSE_USER_GUIDE.md sections rewritten to match the docker/ rework: real layout (context: .., dockerfile: docker/Dockerfile), 4-stage Dockerfile incl. whisper.cpp/MiniCPM-V stages, 4 odek services + 3 sidecars, read_only policy snippets

Verified clean (no changes needed)

PLANNING.md (post-fix), MEMORY.md, EXTENDED_MEMORY.md, SESSIONS.md, STREAMING.md, CACHING.md, API.md (Go SDK surface), WEBUI.md (auth + all numeric limits), SCHEDULES.md, DAILY-WORKER.md, REDACTION_HARDENING.md.

Verification

  • Final stale-identifier grep gate passed (0 hits) before commit; the gate itself caught 3 residual hits (incl. two TestIsPlanMessage citations in PLANNING.md missed by a sub-agent pass) which were fixed pre-commit.
  • Docs-only diff — no Go code touched; CI expected green.
  • Every correction carries file:line evidence from the auditing pass; two auditor false-positives were caught and rejected during triage (sandbox default semantics, max_concurrency default 3).

Maintainer note (code, not docs)

odek init --global writes an explicit "sandbox": false into the template while the CLI treats unset as sandbox-wanted-by-default — worth a follow-up look so fresh configs don't silently opt users out of the default-on posture.

Doc-vs-code audit of all 24 docs/ files + README + AGENTS.md (9 sub-agent
waves, read-only reviewer profile, max 2 concurrent) plus a dedicated
SECURITY.md audit. Every concrete claim (identifiers, flags, defaults,
caps, test names, signatures) checked against source; docs-only changes:

- SECURITY/PLANNING: stale TestIsPlanMessage refs, ToolCallTaints scope,
  guard mcp_schema scope, plan-files bounds row, MCP server-mode
  non_interactive default (read_only), utf16Len claim removed
- CONFIG/CLI: phantom ODEK_TOOL_PROGRESS env, --model default
  (deepseek-v4-flash), repl unknown-flag behavior, serve --no-compaction
- SUBAGENTS/MCP/EXTENSIONS: tools table +4, approval keys hash
  schema+description, per-tool approval for every server, discovery
  fail-fast, +subagent_spawned/completed events
- TELEGRAM/WEBUI/SANDBOXING/README/AGENTS: handler signatures +userID,
  sandbox default-on posture, stale render Start() symbol, event lists
- CHEATSHEET/TOOL_SELECTION/MAINTENANCE/PROVIDERS/DEVELOPMENT: memory
  actions 6->16, phantom skill agent tools removed, artifacts_max_age_hours,
  11 danger classes, docs/ index completed
- DOCKER_COMPOSE_USER_GUIDE: corrected to match the docker/ rework
@cloudflare-workers-and-pages

Copy link
Copy Markdown

Deploying with  Cloudflare Workers  Cloudflare Workers

The latest updates on your project. Learn more about integrating Git with Workers.

Status Name Latest Commit Preview URL Updated (UTC)
✅ Deployment successful!
View logs
odek 7e3fa8e Commit Preview URL

Branch Preview URL
Aug 31 2026, 10:51 AM

@jkyberneees
jkyberneees merged commit f3a9d39 into main Aug 31, 2026
10 checks passed
@jkyberneees
jkyberneees deleted the fix/security-docs-accuracy branch August 31, 2026 10:56
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant