pi-ai resolves a request's apiKey override only through a provider that
declares an api-key method: resolveProviderAuth short-circuits to that
method when the override is present, and otherwise falls through to the
credential store and then to ambient discovery. A provider with no
api-key method at all therefore resolves to nothing, and the request
fails with "Provider is not configured" before any network I/O.
Two routes hit that. openai-codex ships OAuth alone, so moving off the
/compat dispatch broke a profile that names a key for it — the old path
handed the token straight to the provider. And a catalog route naming an
api was being rebuilt with the harness's own auth, so `openai: {api:
openai-completions}` stopped reading OPENAI_API_KEY, contradicting the
documented promise that omitting a credential keeps provider-native
discovery.
Auth is now one decision for both constructions. A catalog route keeps
its installed provider's auth, through an api override too: which
environment a provider reads belongs to the provider, not to the wire
format its models speak. A catalog provider with no api-key method gets
the harness method beside its own, but only when the profile names a
credential — a keyless codex profile keeps the honest refusal, since
this adapter holds no OAuth store to resolve through.
Materialization now spreads the installed entry instead of enumerating
the result, so a Model field this package does not model survives a
pi-ai upgrade; headers went missing from an nvidia route exactly that
way once already. providerInfo reports the configured displayName, which
also joins the registration facts so a rename re-registers rather than
leaving the old label in every selector. A refused registration swap
gets its own diagnostic naming the route, matching the directory swap
beside it.
The README documented endpoint interrogation this layer does not
implement, and still described unknown providers as kept-last-good after
they became legal declarations refused at the write point. The Agent
Note claimed per-model reasoning configurability the schema never had,
required capacities the route now defaults, and stated an apiKey
override that short-circuits unconditionally.
The adapter omits the seam's reasoning field whenever a model carries no
reasoning metadata, which is the model's own property and says nothing
about where the model came from. Both the JSDoc and the Agent Note read
as though only hand-declared models were meant, so a reader would infer
that the 251 installed-catalog models pi-ai marks as non-reasoning still
offer their single off level. They do not, and that is the point: a
picker holding only off misrepresents a provider that thinks by default,
because off dispatches the same bytes as naming no effort at all.
Behavior is unchanged; only the prose that describes it was narrower
than the contract. adapter.spec.ts already pins the catalog case through
openai/gpt-4.1 and catalog.spec.ts pins the hand-declared one.
Three defects surfaced while driving the Models page.
A hand-declared model needed an explicit contextWindow and maxTokens,
but a provider listing usually returns ids and nothing else — so the
page happily wrote a profile the adapter then rejected, which took the
whole namespace down silently. Capacities now fall back to the route's
`defaultContextWindow` (262,144) and `defaultMaxTokens` (32,768). Both
are guesses by construction, which is why they are route fields a
deployment corrects once rather than constants buried in the adapter;
the fallback sizes the model and never becomes a per-request cap.
That silent failure was the second defect. A schema-valid profile the
adapter could not serve was stored and only rejected later, disabling
every route in the namespace with nothing said. `dsh-settings` gains an
optional `validate` on registration — a check for what a schema cannot
express — and `llm-pi-ai` refuses an unserviceable section at the write
that produced it. A stored section that fails keeps the namespace's last
good value, as a schema failure already did, so an externally edited
document still cannot strand the owner. The plugin's own last-good
fallback goes with it: nothing reaching it can fail any more.
Third, a model with no reasoning metadata advertised the single level
`off`, which pi-ai translates to *omitting* the reasoning option — the
same request naming no effort produces. Selecting it disabled nothing,
so a provider whose default is to think kept thinking with `off` shown
as selected. Such a model now reports no reasoning capability at all,
which is the seam's way of saying the control is unavailable, and the
per-model `reasoning` flag is gone: without a thinkingLevelMap to spell
levels it could only invent them.
The protocol table narrows to the three a hand-declared route reaches
today, most-reached first so a surface offering a choice defaults to the
one gateways actually speak.
Review found four defects in the declared-provider work.
`PiAiAdapter` reused one `Models` collection and mutated it whenever the
configuration changed. `Models.streamSimple()` resolves its provider
lazily — when the stream is first consumed, which is after the adapter
awaits the route's credential — so a configuration change landing in
that window let an in-flight request finish under a configuration it
never resolved against, or fail on a provider that no longer existed.
Each resolution now produces an immutable snapshot and every operation
captures one before its first await, which is what makes the seam's
per-step freeze (`llm.prepareCall()`) hold end to end: switching models
mid-reply takes effect on the next step, never inside the one in flight.
`defaultMaxTokens` was materialized from the catalog's `Model.maxTokens`.
The two answer different questions: pi-ai requires that field as the
model's output capability, while the seam's is a cap the deployment
chose to send on requests naming none, so every request had started
carrying a number nobody picked. Only an explicitly configured cap
reaches the seam now.
The configurable-provider directory was refreshed by disposing its
registration and making a new one. A candidate set the registry refuses
— a profile keyed `deepseek-official`, which llm-deepseek declares —
left the whole directory withdrawn and the Models page empty, silently,
because the settings callback contains the failure. The seam's
registration handle now carries `replace()` with the same
validate-first atomicity `registerAdapter` has.
The protocol table offered every pi-ai streaming API, including four
whose authentication a profile cannot express: Bedrock signs with SigV4
over AWS credentials and a region, Vertex needs a project, a location,
and ADC, Azure needs provider environment plus an api-version, and Codex
uses OAuth. Offering them handed back routes that cannot authenticate.
Catalog routes still reach them through their own provider.
A pi-ai route had to name an installed catalog provider, served that
catalog's models verbatim, and could override only the endpoint. An
OpenAI-compatible gateway, a self-hosted server, or a model newer than
the pinned pi-ai release was therefore unreachable, and a stale context
window could not be corrected without upgrading the package.
A route is now a declaration whose defaults come from the installed
catalog. `catalog.ts` merges that catalog under the profile's own model
entries, `provider.ts` builds the pi-ai Provider (reusing the catalog
provider when the route keeps its protocol, so implementations this
package cannot reconstruct keep working), and the adapter serves every
operation from one `createModels()` collection. That also retires the
`@earendil-works/pi-ai/compat` import, which pi-ai documents as a
temporary entry point it deletes with its ModelManager migration.
Credentials stay on the harness seam: the resolved key rides the request
as pi-ai's highest-priority auth override, so `Models` holds no
credential store and a named-but-missing reference still fails loud
instead of falling back to an unrelated ambient key.
A model's configured maxTokens now reaches the seam as defaultMaxTokens.
The merge staged the `--ours` consistency records before the merged prose was
re-recorded, so the committed hashes still describe the pre-merge content and
`verify-translation-pairing` rejects the pair. Re-record all three against
what actually merged.
- `relay` resolves its sender in `contextBody` like every other form. It was
the one shape whose marker could claim a form the body did not render: an
unreadable sender fell back inside the body while the row still said relay,
contradicting the contract this PR's own note states.
- `recall` requires the retained, omitted, and truncated fields. Completeness
is what the card exists to report, so a reference that cannot state it is
not a readable recall — showing the label alone presents a confident card
over unknown loss.
- The snapshot body states the supersession its producer framing line carries.
That line is the one part of the model-facing text no section contains, and
unlike an instruction context's `<system-reminder>` it states the form's own
semantics rather than wrapping content.
- `GoalMessageSource` is a discriminated pair, so `{ form: 'notice' }` without
its account no longer compiles. The guarantee this PR claims now holds at
that seam too, not only through `ContextFormed` on plugin sources.
- Goal and tool-goal summaries are bounded by a shared `boundContextSummary`,
which tool-tasks now uses as well. A goal objective is unbounded caller text
in exactly the way a task label is.
- The runtime snapshot interpolates once per request: agent-loop renders the
sections and joins them through `joinContextSections`.
- Every form's fallback branch is pinned, not only the notice one.
Four values complete the vocabulary, so the opaque body is reached only by
producers that genuinely promise no shape.
`snapshot` — current state a later snapshot supersedes. system-prompt now
exposes `renderContextSections()`, the named contributions
`renderContextSnapshot()` already joins for the model, so the body attributes
each part to the subsystem that produced it instead of re-splitting joined
prose. The runtime snapshot, time-context, and tmux-context declare it.
`notice` — a one-off account of what just happened, declared by tool-tasks,
goal state changes, tool-goal wrap-up, plan-mode switches, and
repeat-tool-guard. Its `summary` rides the COLLAPSED row: these five are the
majority of shipped producers and none of them needs expanding to be read.
The task summary bounds itself because its inputs are unbounded caller text.
`relay` — a message another agent addressed to this one; both subagent
sources declare it and the body names the sender above what it said.
`recall` — material lifted from another session's log. session-reference
needed no new field: its references already record retained and omitted
counts and the truncation flag, which the body shows first, because recalled
context is bounded on the way in.
`ContextFormed` is now discriminated by `form`, so a producer cannot declare
a shape without the facts that shape is presented from — a notice without its
summary, or a snapshot without its sections, fails to compile.
Only the two hook bridges stay opaque, by design: their content is whatever
an external program printed, so no shape can be promised for it. Unknown
kinds and unreadable records land there too.
The composer ring, percentage, and `~used / capacity` header read
`contextPressure.pressureTokens`, which moves only when a request reports
usage. Compaction reports none — compact-basic summarizes through a direct
`ctx.llm.stream()` call and appends only its own `compact/*` records plus the
replacement `user/message` — so the meter was frozen across the one action
taken to change it. Driving a real `compactNow` through the agent loop:
BEFORE compact: ring=4% header=~4227/100000 rows=[18, 0, 4365]
AFTER compact: ring=4% header=~4227/100000 rows=[18, 0, 286]
The composition rows fell 93%; the ring did not move, and would not until an
entire further turn completed. The panel then contradicted itself by more than
an order of magnitude at exactly the moment a reader opens it.
`contextPressure` now also publishes `projectedTokens`: the provider sample
plus the heuristic repricing of everything the surface gained or lost since
that sample, clamped at zero, folded through the shared `surface-fold.ts`. The
sample is stamped before the same event joins the surface, so an
`assistant/message` anchors against the surface its own request carried. Only
the delta is estimated, so the figure stays provider-anchored — the estimator's
CJK and JSON-schema underpricing stays out of the occupancy number — while
reacting the moment content lands or a span is shadowed. Same run after:
BEFORE compact: ring=4% header=~4323/100000 (pressure=4227, projected=4323)
AFTER compact: ring=0% header=~ 244/100000 (pressure=4227, projected= 244)
`contextOccupancy` prefers the projected figure and falls back to the bare
sample, so a projection restored from a pre-field checkpoint degrades to the
old behavior rather than disappearing. `stateVersion` moves to 3.
`contextBreakdown.messageTokens` and `measure().surfaceTokens` answer the same
question in the same heuristic vocabulary, and the panel's composition rows are
only honest while they agree. Each owner carried its own copy of the positional
fold — same pricing, same `{seq, tokens}` node list, same replace-range lookup
and guard, differing only in mutable versus immutable application — so an edit
to either one would have moved the panel away from `measure()` with both sides
still green. The duplication gate caught the shared 62 tokens.
`src/surface-fold.ts` now owns `foldSurfaceTokens`: total, allocation-fresh,
returning the event's price, the next surface, and the signed total delta. The
service assigns that result where it used to prepare a commit closure, which
keeps its validate-before-mutate replay transaction intact — the fold throws
before any state is touched, so a malformed event still fails identically on
every retry. `_prepareSurfaceMutation` and `_estimateSurfaceEvent` go away with
it, and the projection's apply drops to one call.
Covers the identity with a session that appends and then compacts, asserting
the projection figure equals the service surface at each boundary; the test
fails when either side of the fold is perturbed.
Assistant footers and the stats line gain TTFT/tok-per-second readings
folded from step timings; context occupancy moves off the stats line onto
a composer ring whose panel shows a heuristic system/tools/messages
breakdown from the new token-meter contextBreakdown session projection.
The e2e store and the web-search store still named a `.env` path; the
e2e one also wrote dotenv syntax, which the YAML document rejects. That
path now names the ordinary environment layer, so a test pointing the
credential store at it asserts the distinction this PR removes.
Injected context reached the transcript as one anonymous shape whose
expanded body was the whole message serialized as inline JSON, so the
escaping collapsed the only readable part — the model-facing prose —
into a single line.
`MessageSource` gains an optional producer-declared `form`, a small
semantic vocabulary of information shapes independent of `kind`: kind
says who produced the context, form says what shape it is, so several
producers may share one presentation. Two values ship.
`instructions` (workspace-context) lists the reconciled files above
their text and keeps the `<system-reminder>` framing verbatim, because
the framing is part of what the model read. `catalog` moves dsh-tool-skill
off the shared plugin kind onto a `skill-catalog` source carrying the
published name/description entries, and the body lists those instead of
re-parsing `<available_skills>` out of the prose. Catalog identity moves
with it: the republish digest now covers the durable entries, deleting
the text-slicing that recovered them from a logged message.
Everything else renders the opaque body — the model-facing text with its
real line breaks, then the remaining provenance as fields. That is the
documented default, not a leftover: a resumed, forked, or foreign log
must render whether or not its producer is mounted here, which is why
the classification lives in the durable source rather than a client-side
table keyed by producer.
Two had real security consequences:
The bootstrap rejection ran on npm dotenv's parser while process.loadEnvFile
applied the file with Node's own. Two independently maintained dialects meant
the check and the thing it guards could disagree: a name Node accepts but the
checker misses would reach process.env unchecked, and BASH_ENV there runs a
file of the project's choosing on every `bash -c` the bash tool issues. Parse
once with node:util's parseEnv — the same engine loadEnvFile uses — and assign
the entries already checked, which also drops the dotenv dependency.
llm-pi-ai still returned a literal profile.apiKey ahead of everything, and it
registers a settings namespace, so the defect removed from llm-deepseek
survived intact in its design twin. The field is gone from the profile schema,
the resolution path, and the tests.
The rest are consistency and documentation defects the review named:
- verify-config-source-ownership did not scan the Python runtime's bundled
cordis.yml, which still inlined apiKey and baseURL. Both are covered now, and
the line-anchored INLINE_DENY documents that it is a tripwire, not a parser.
- The deny list missed NODE_TLS_REJECT_UNAUTHORIZED, the askpass hooks, the
GIT_CONFIG_* redirections, and PYTHONHOME — all implied by its own stated
rule about what a variable does.
- Snapshot lookups folded case on Windows, where environment names are
case-insensitive and an exact-match Map could miss a higher-ranked layer.
- The credentials note claimed a read-time permission check was "not taken"
while this PR implemented it; the credentials-local README still described
two layers, live process.env reads, dotenv-era limitations, and a renamed
anchor; the llm-deepseek README still advertised the removed literal apiKey;
and web.ts and base.cordis.yml kept personal-overlay wording.
- The ownership note's literal-apiKey claim now names its scope: the
web-search providers keep a literal field but register no settings
namespace, so nothing can shadow a stored credential through them.
Review found five real defects in the configuration-source work, all confirmed
against the code rather than argued:
1. The note claimed --config outranks settings.yaml. It does not: the settings
seam registers a plugin's cordis entry config as the `base` layer and the
user section layers over it, and the seam cannot tell a shipped value from a
--config one. The note now states shipped reality and names --config-replace
as the lever for a deployment that must win. Separately, a literal `apiKey`
in settings outranked both the environment and .credentials.yaml — the field
is removed, so configuration carries a reference and nothing else.
2. DEEPSEEK_SEARCH_BASE_URL was functionally deleted: the shipped inline went
away without the provider learning to read it. It now resolves from the
environment snapshot, as the README always claimed.
3. The bootstrap deny list missed the interpreter start-up hooks. BASH_ENV is
the sharpest: `bash -c` sources it on every bash tool call, so a project
.env could run a file of its choosing before every command. The list now
covers BASH_ENV and its per-language siblings, the Git hook commands, and
the remaining preload and CA variables, organised by what a variable does
rather than which runtime owns it.
4. YAML parse errors quoted the offending source line — which in a credentials
document is the secret — into boot stderr and the watcher's logger. Only the
error code and position are reported now, in credentials-local and
settings-local alike, pinned by a test that asserts the secret is absent.
5. 0600 governed only files the harness wrote. A hand-created 0644 document was
read normally. POSIX now checks the mode before reading contents, at boot
and on every reload; Windows has no mode to inspect and is skipped rather
than faked.
The project a session is launched in is trusted by default, with no prompt and
no stored trust record: it may supply its own endpoint, ordinary variables, and
a key ranked below the managed store. Trust stops at the harness itself — a
discovered file still cannot set DSH_PERMISSION_MODE, PATH, BASH_ENV, or the
rest, because those take effect with no user action, before any turn, outside
the permission policy and the sandbox.
$DSH_HOME/.env had just become an ordinary environment layer, which left the
harness resolving user-facing values from a flattened process.env that could
no longer say where a value came from. A key stored through the web page
stayed shadowed by an older key in the user's own .env. An endpoint could be
redirected by the project: the invoking directory's .env is materialized like
every other layer, and a base URL decides where a resolved API key is sent, so
a DEEPSEEK_BASE_URL written into a model-editable workspace would send the
user's credential — and the prompts carrying their code — to whatever host
that file named.
Give every user-facing value one ordering, with four kinds of source:
explicit for this run per-operation override, CLI argument
> authored by deployment --config / --config-replace
> this launch's shell inherited process environment
> product-managed store settings.yaml, .credentials.yaml
> discovered file $DSH_HOME/.env
> defaults schema default, shipped base, public default
The domains differ only in which tiers exist. The earlier split — credentials
ranking the environment over the managed file while settings ranked over the
environment — was inconsistent: the distinguishing fact is who authored the
source, not the domain.
packages/util/environment owns an immutable snapshot with per-layer
provenance. getFrom(name, sources) searches only the layers a caller names,
and omitting one is a refusal rather than a demotion: the adapters ask for
['process', 'user-env'], so no reordering can let a project file back into a
decision it was excluded from.
isBootstrapOnly rejects, before anything is materialized, any .env setting a
variable that governs how a process launches (PATH, SHELL, NODE_OPTIONS,
LD_PRELOAD), where code or model-visible instructions load from (the whole
DSH_* namespace, HOME, XDG_*), or how the network is reached (proxy and CA
variables). The namespace is denied wholesale so a switch added later cannot
become settable by being forgotten, and there is no opt-out.
verify-config-source-ownership keeps both rules: no unregistered process.env
read under packages/*/*/src (26 allowlisted with reasons), and no apiKey,
baseURL, or headers inlined from the environment in shipped Cordis config —
removing those inlines is what makes the deployment tier meaningful.
A catalog entry may now carry its own `maxTokens`, matching the shape the
pi-ai adapter already exposes. Exact-model resolution prefers it over the
profile value, so capping one model no longer means capping the route.
Defaults are unchanged: an entry without a cap, and any unlisted
pass-through id, still resolve to the profile `maxTokens` (256,000), and the
shipped catalog keeps its context windows.
$DSH_HOME/.env carried two incompatible jobs. As credentials-local's writable
secret store it could not be hoisted into process.env — hoisting makes every
stored key read as a read-only launch override and blocks rotation from the
TUI and the web page. But its name and dotenv format promise an environment
file, so a DEEPSEEK_BASE_URL sitting beside a working DEEPSEEK_API_KEY in the
same file was silently ignored: only the credential provider read the
document, and it addresses credential references alone.
Split the two jobs into two files.
.credentials.yaml is the provider-managed store: a strict YAML mapping of
CredentialRef to non-empty string, no version field, no wrapper level. Because
it holds credentials and nothing else, a non-mapping root, a non-identifier
key, a non-string value, an empty string, a duplicate key, and malformed YAML
are all rejections rather than skipped entries — loud at boot and at a write,
warn-and-keep-last-good on a live reload. The dotenv physical-line editor
gives way to a patch of the parsed document, so comments and untouched entries
keep their formatting and any string value round-trips, multi-line included.
Writer lock, read-modify-write, atomic 0600 write under a 0700 directory,
watcher, self-write suppression, and quiescent disposal are unchanged.
$DSH_HOME/.env becomes the user's ordinary environment layer. app-boot's new
loadLayeredEnv loads the invoking directory's .env then the Harness home's,
giving user < project < inherited; the home resolves from the inherited
environment first, so a project .env cannot redirect it.
Credential precedence is unchanged: the live environment still wins read-only
over the file, and shadowed writes still reject. Whether a provider-managed
store should instead win over the environment is a separate decision.
No migration: a key already in $DSH_HOME/.env keeps resolving through the new
environment layer, as a read-only env source that shadows the stored one.
TurnEndReasonMap.error now carries a single `error: LlmFailure` field:
an LlmError keeps its structured facts, any other error flattens to
errorChain text under the UNKNOWN code. Consumers read message/code
directly instead of defending against an unknown union — this also fixes
errorChain() rendering structured failures as '[object Object]' in the
TUI and ACP error paths. Document the turn-stopping contract: a
concludesTurn result never short-circuits already-submitted next-step
work (same-step additionalContexts or racing steering still runs), data
decides.
