Wiki and learning

Two things learn in codeloop. The wiki learns facts about the codebase: gotchas, decisions, concepts. The lanes learn how work should flow: a gate that keeps rejecting gets a tighter check. Both go through a person before anything is enforced.

Wiki pages

A page is a markdown file under .codeloop/wiki/<kind>/<slug>.md with a title, a kind, a scope of file globs, a frequency and a severity.

codeloop wiki capture \
  --title "CSV export must quote commas" \
  --scope "src/export*" \
  --body "A value with a comma breaks the row unless quoted." \
  --kind gotcha --card c-001

  warning  freq 1   CSV export must quote commas  .codeloop/wiki/gotchas/csv-export-must-quote-commas.md

Capturing the same title again does not make a second page. It raises the frequency:

  warning  freq 2   CSV export must quote commas  .codeloop/wiki/gotchas/csv-export-must-quote-commas.md

Inject before a stage

Every rendered host file tells the agent to run wiki inject before touching files. It prints the pages whose scope matches:

codeloop wiki inject --files src/export.ts

## CSV export must quote commas (gotcha, freq 2, warning)

A value with a comma breaks the row unless quoted.

The stage brief from codeloop brief includes the same pages for the files the stage names. wiki list prints every page; wiki lint reports broken links (exit 1), stale pages and duplicate titles.

Frequency raises severity

codeloop learn

learn applies two thresholds from config.yaml:

With one page at frequency 2 the output is nothing crossed a threshold. A build-lane stage that puts codeloop check gotchas --files ... in its check turns a critical page into a blocking check, not a reminder.

Competitor pages

codeloop wiki competitor keeps one page per competitor with its links at the top. The scan lane reads the changelog link on each page every Monday and proposes a plan card when something new appears. The build lane's research stage copies its "how others show it" row onto the competitor's page when the stage passes. The founder's week shows both.

Lanes learn from rejections

Rejections at a gate and stuck stages are counted per lane version. At three of either on one stage, codeloop lane propose writes a proposal folder:

.codeloop/proposals/market-draft-1/
  lane.yaml   the lane with version +1 and the rejection notes added to the stage's notes:
  why.md      which cards, when, and each note, in a table
  eval.yaml   fixtures to fill in

The proposal is a change to the lane file, not to a prompt. The notes land in every future brief for that stage. If the check should be stricter, the owner edits lane.yaml and adds a fixture to eval.yaml that the new check must fail:

fixtures:
  - name: claim-without-proof
    setup: "printf 'claim: fast\\n' > marketing/c-1/blog.md"
    expect: fail
  - name: claim-with-proof
    setup: "printf 'claim: fast\\nproof: 0.4s\\n' > marketing/c-1/blog.md"
    expect: pass

codeloop lane eval market-draft-1

  ok   draft / claim-without-proof: expected fail, got fail
  ok   draft / claim-with-proof: expected pass, got pass
  ok   replay c-009 / draft
  eval green for market-draft-1

eval runs each fixture in an empty temp dir against the changed check, and replays the lane's finished cards through it so a stricter check does not fail work that was already approved. A changed check with no expect: fail fixture is refused as vacuous. promote is refused before a green eval, and an agent cannot promote at all.

codeloop lane promote market-draft-1

The lane is now version 2. Version 1 is kept, and codeloop lane rollback market restores it. codeloop stats --lane market --compare v1 v2 shows whether the change helped.

What is learned where