Skip to content

docs: keep only what describes the system as it is, and generate the rest - #55

Merged
rldyourmnd merged 1 commit into
mainfrom
docs/current-state-only
Aug 29, 2026
Merged

docs: keep only what describes the system as it is, and generate the rest#55
rldyourmnd merged 1 commit into
mainfrom
docs/current-state-only

Conversation

@rldyourmnd

Copy link
Copy Markdown
Contributor

Three decisions applied: documentation is written for agents, anything derivable from code is generated rather than typed, and the tree carries only current state because git carries the history.

Removed — 33 files, 11,884 lines

29 completed-phase evidence and migration-plan documents, two superseded ADRs (0001, 0017), two architecture documents describing the pre-migration design. Git holds all of it.

gds-completion-plan.md was not purely historical. Three runbooks cited it as their authority, and it carried the canonical-owner table, the status vocabulary and the change protocol — none of which are migration steps. Those are extracted into docs/contracts/authority-and-change-protocol-v1.md; the runbooks point there. The migration order itself is gone.

Generation instead of prose

docs/contracts/harness-adapters-v1.md claimed:

scripts/validate_harness_docs.py fails when this list, the registry, and that constant disagree, so the three cannot drift apart silently as they did before.

That script does not exist. The document described a guard nobody had built, and the drift it promised to prevent is exactly what happened — the doc listed seventeen identities while the code had seven.

So the list is now generated, not validated. scripts/generate_harness_docs.py:

  • renders the identity blocks in the contract and harnesses/README.md from the registry;
  • refuses to run when the registry and core/harness.CanonicalIDs disagree;
  • --check runs in validate_ci_tier.sh beside the ADR index check.

Proved it fails in both directions before relying on it — injecting a stale id into the README, and adding an id to CanonicalIDs, each exit 1.

Corrected statements that had become false

  • README: "all seventeen harnesses", and example-org/example-harnesses as evidence owner (now NDDev-it-com/setup-systems)
  • architecture README: binary list still naming the deleted gds-zcode-runtime-driver
  • adapter table: rows for three removed harnesses
  • ADRs 0035 and 0036 carried consequences that are now wrong — annotated in place with what changed rather than rewritten, because the decision was real when it was taken

Verification

  • go test ./... — 64 packages, zero failures
  • scripts/validate_ci_tier.sh fast — exit 0
  • generate_harness_docs.py --check — clean
  • no dangling reference to any deleted file remains

https://claude.ai/code/session_01GFeam4oG1UnwF9r55pXEyp

…rest

Three decisions applied: documentation is written for agents, anything
derivable from code is generated rather than typed, and the tree carries only
current state because git carries the history.

Removed 33 files: the 29 completed-phase evidence and migration-plan documents,
the two superseded ADRs, and the two architecture documents that described the
pre-migration design. Nothing is lost -- git holds all of it -- and the tree
now shows the system as it is rather than as a record of how it got here.

`gds-completion-plan.md` was not purely historical: three runbooks cited it as
their authority, and it carried the canonical-owner table, the status
vocabulary and the change protocol, none of which are migration steps. Those
are extracted into docs/contracts/authority-and-change-protocol-v1.md and the
runbooks point there. The migration order itself is gone.

Generation instead of prose. `docs/contracts/harness-adapters-v1.md` claimed
that `scripts/validate_harness_docs.py` "fails when this list, the registry,
and that constant disagree, so the three cannot drift apart silently as they
did before". That script did not exist. The document described a guard nobody
had built, and the drift it promised to prevent is exactly what happened: the
doc listed seventeen identities while the code had seven.

So the list is generated now, not validated. `scripts/generate_harness_docs.py`
renders the identity blocks from the registry, refuses to run when the registry
and `core/harness.CanonicalIDs` disagree, and `--check` runs in
`validate_ci_tier.sh` beside the ADR index check. Proved it fails in both
directions before relying on it: injecting a stale id into the README, and
adding an id to CanonicalIDs, each exit 1.

Corrected statements that had become false: the README's seventeen harnesses
and its `example-org/example-harnesses` evidence owner, the architecture
binary list still naming the deleted zcode driver, and the adapter table's
rows for three removed harnesses. Two accepted ADRs carried consequences that
are now wrong; those are annotated in place with what changed rather than
rewritten, because the decision was real when it was taken.

Verified: go test ./... 64 packages zero failures; validate_ci_tier.sh fast
exits 0; generate_harness_docs.py --check clean; no dangling reference to any
deleted file remains.

Claude-Session: https://claude.ai/code/session_01GFeam4oG1UnwF9r55pXEyp
@rldyourmnd
rldyourmnd merged commit 7e7b205 into main Aug 29, 2026
7 checks passed
@rldyourmnd
rldyourmnd deleted the docs/current-state-only branch August 29, 2026 06:39
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