[roodboi]

Linear: HACK-1158 (WU27-HYDRATE). Stacked on #340 (wf/f7-stable-ciphertext-b16), which provides semantic ciphertext reuse and the rename-committed layer write. Review and merge #340 first.

Final behavior

hack env hydrate imports many values into the machine-local env layer in one commit:

<secret-source> | hack env hydrate --secret
hack env hydrate --secret --file .secrets.env
hack env hydrate --secret --file secrets.json --format json --env qa --service api

• Input never in argv. Entries come from piped stdin (a TTY is refused before reading) or --file (regular files and pipes such as <(producer)). Input is bounded at 1 MiB. The env format is strict (KEY=value, comments, export, 'literal', "escaped"). --format json takes one object of non-empty strings. Repeated keys are refused in both formats; JSON input is not resolved last-wins. One malformed entry refuses the whole input. Messages name at most a line number.
• Only the machine-local layer is written (hack.env.local.yaml or hack.env.<overlay>.local.yaml). Tracked defaults are untouched. A git-tracked local layer is refused. Inside a git checkout where git cannot say whether the layer is tracked (exit 128 from a safe.directory refusal or a damaged index, or git missing), hydration also refuses, so the tracked check fails closed. Outside any checkout it proceeds.
• Idempotent. An entry that already holds its value is kept: a secret keeps its ciphertext (F7's semantic reuse) and a plain scalar keeps its YAML type. When nothing changes, the file is not rewritten, so hand edits and comments survive and Git shows no diff. A changed value gets new ciphertext with a fresh IV. A repeat must use the same flags: a plain import over a key the layer stores as a secret is refused (secret_downgrade) rather than silently replacing ciphertext with plaintext.
• Local overrides preserved. Local entries outside the input, other scopes, and removal markers for other keys are kept. A key in the input replaces that key's local entry.
• All-or-nothing. The whole layer is published by one rename (F7's atomic write). An interrupted import leaves every key at its previous value.
• Refusals without partial state (closed, value-free, ClosedCliError):
  • tracked_layer: git tracks the target local layer.
  • tracking_unknown: inside a git checkout, git could not report whether the layer is tracked.
  • unreadable_layer: the local layer exists but is not a valid env file (invalid YAML, wrong shape or version). It is never replaced.
  • wrong_key: a secret already in the layer does not decrypt under the reader key, or no key exists. The key file is checked before HACK_ENV_SECRET_KEY, as readers do. No key is minted.
  • key_env_only: --secret with only HACK_ENV_SECRET_KEY and no key file, because minting a file would replace that key for every layer.
  • secret_downgrade: a plain import of a key stored as a secret in the target scope.
  • locked / stale_lock: hydrations are serialized by .hack/.internal/env-hydrate.lock (exclusive create, holds the PID). A lock from an exited process is reported with its path, never taken over.
  • concurrent_write: compare-and-swap. The layer must still equal the bytes read before staging and again after fsync, just before the rename.
• Values never printed. Text and --json output list only added, updated and unchanged keys.
• Cross-command hardening. Unexpected arguments now reports a count instead of echoing the arguments, so hack env hydrate API_TOKEN=value cannot print the value. This applies to every command.

Docs: docs/env.md gets a hydration section that lists every refusal, plus a matrix row. docs/reference/cli.md is regenerated. Pattern card HACK-P08 lists hack env hydrate --file <path> --secret and its remaining gap, re-rendered into plugins/hack/skills/hack-cli/references/patterns.md.

Rework after verification (head fba3e87c)

• secret-scan (blocker): 57a82c2, pushed separately, builds the synthetic HACK_ENV_SECRET_KEY at runtime and adds the fingerprint of the old line in 06f2623 to .gitleaksignore. CI scans the PR's full history, so the runtime change alone would not have passed. fba3e87 moves that test onto a shared runtime-built key helper.
• Tests now pin the unreadable_layer refusal (invalid YAML, a non-mapping scope, a wrong version, in secret and plain mode, bytes unchanged) and the reader-key priority (key file plus a different HACK_ENV_SECRET_KEY: plain and secret hydration succeed, and readers open every entry).
• Fail-closed tracked check: new tracking_unknown refusal, described above.
• Secret downgrade: new secret_downgrade refusal, described above.
• JSON duplicate keys: refused with duplicate_key, including escaped spellings such as API_TOKEN. Nested keys and quotes, commas or brackets inside values are not miscounted.
• Docs: the refusal list in docs/env.md is complete, including unreadable_layer and stale_lock.

Validation (local, macOS host, head fba3e87c)

• tests/project-env-hydrate.test.ts, tests/env-hydrate-input.test.ts, tests/cli-command.test.ts, tests/integration-ownership.test.ts, tests/cli-spawn-env-isolation.test.ts: 48 pass, 0 fail.
• Broader set of 16 files (the above, plus agent-instruction-source, cli-reference-docs, env-command-matrix, env-key-argument-redaction-cli, env-value-stdin(-cli), pattern-catalogue, project-env-config, local-inheritance, stable-ciphertext, worktree-key): 204 pass, 0 fail. The run before it was at load average ~68 on the shared host and had 3 five-second timeouts in tests unrelated to this change (for example a whitespace-only key file is refused). The rerun passed.
• bun run --cwd packages/cli typecheck: clean. bun run --cwd packages/cli check: 0 errors (34 existing warnings). Biome on every changed file: clean.
• gitleaks 8.30.1 with the repo config and CI's --log-opts='--full-history --diff-filter=tuxdb HEAD': 915 commits, no leaks.
• No CLI surface or guidance change in the rework, so the CLI reference and plugin renders are unchanged.

Negative controls

Every new test was proven with a mutation. Mutants killed:

• unreadable_layer guard removed
• reader-key priority swapped (environmentKey ?? fileKey)
• tracking_unknown check removed
• checkout marker always reported as present (8 tests fail)
• secret_downgrade guard removed
• JSON duplicate check removed

JSON scanner mutants, also all killed:

• root does not expect a key
• a comma at any depth expects a key
• any opening brace expects a key
• a key is not consumed
• escapes are ignored
• keys are compared without decoding
• the scan does not skip to the string's end

Two mutants survived and were found to be equivalent: a redundant depth check and a nested-open reset. Both were removed from the code. The original change's 17 mutants were not rerun; the code they target did not change.

A secret-scan negative control used the base branch's config with no fingerprint: the PR range still reports 1 leak, from 06f2623. The current test file scans clean on its own.

Tests use synthetic values, private HOME via privateCliEnv, and temp git repos. The tracking_unknown fixture corrupts the temp repo's own .git/index. No test touches Docker, the keychain or real config.

For the reviewer

• Acceptance interpretation: the issue says "overrides survive". Local entries outside the input survive, but an imported key replaces an operator's deliberate --local override of that same key. This is disclosed in the HACK-P08 missingContract and in the docs, and the pattern card stays partial. Please confirm whether this meets the bar.
• secret_downgrade is a new refusal with no override flag. The way out is --secret, or hack env unset <key> --local and then a plain import. hack env add without --secret still downgrades, as before.
• tracking_unknown refuses on a machine without git when the project has a .git entry. This is deliberate: fail closed.

Remaining gaps

• Hydration cannot tell an operator's deliberate local override of an imported key from a stale hydrated value. An imported key always replaces its local entry.
• hack env set --local does not take the hydration lock. Only the pre-rename recheck guards against it, which narrows the window but does not close it.
• A SIGKILLed hydration leaves a stale lock (reported with its path) and a staging file. Both are ignored by git, and the lock needs manual removal. SIGINT/SIGTERM of the real CLI while it holds the lock is not tested; only SIGKILL is.
• --env omitted writes the base local layer, as hack env add does today, not the configured default overlay.
• Host-bound gaps from the issue are not covered here: native credential approval, runtime-injection receipts on a real stack, and log inspection on a live runtime. No production hydration was run. Hosted CI on fba3e87c is pending.

🤖 Generated with Claude Code